Commit
•
3f82a19
1
Parent(s):
708abd6
Upload 4 files
Browse files- chatfuncs/__init__.py +0 -0
- chatfuncs/chatfuncs.py +1032 -0
- chatfuncs/ingest.py +655 -0
- chatfuncs/ingest_borough_plan.py +14 -0
chatfuncs/__init__.py
ADDED
File without changes
|
chatfuncs/chatfuncs.py
ADDED
@@ -0,0 +1,1032 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
import os
|
3 |
+
import datetime
|
4 |
+
from typing import TypeVar, Dict, List, Tuple
|
5 |
+
import time
|
6 |
+
from itertools import compress
|
7 |
+
import pandas as pd
|
8 |
+
import numpy as np
|
9 |
+
|
10 |
+
# Model packages
|
11 |
+
import torch.cuda
|
12 |
+
from threading import Thread
|
13 |
+
from transformers import pipeline, TextIteratorStreamer
|
14 |
+
|
15 |
+
# Alternative model sources
|
16 |
+
#from dataclasses import asdict, dataclass
|
17 |
+
|
18 |
+
# Langchain functions
|
19 |
+
from langchain.prompts import PromptTemplate
|
20 |
+
from langchain.vectorstores import FAISS
|
21 |
+
from langchain.retrievers import SVMRetriever
|
22 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
23 |
+
from langchain.docstore.document import Document
|
24 |
+
|
25 |
+
# For keyword extraction (not currently used)
|
26 |
+
#import nltk
|
27 |
+
#nltk.download('wordnet')
|
28 |
+
from nltk.corpus import stopwords
|
29 |
+
from nltk.tokenize import RegexpTokenizer
|
30 |
+
from nltk.stem import WordNetLemmatizer
|
31 |
+
from keybert import KeyBERT
|
32 |
+
|
33 |
+
# For Name Entity Recognition model
|
34 |
+
#from span_marker import SpanMarkerModel # Not currently used
|
35 |
+
|
36 |
+
# For BM25 retrieval
|
37 |
+
from gensim.corpora import Dictionary
|
38 |
+
from gensim.models import TfidfModel, OkapiBM25Model
|
39 |
+
from gensim.similarities import SparseMatrixSimilarity
|
40 |
+
|
41 |
+
import gradio as gr
|
42 |
+
|
43 |
+
torch.cuda.empty_cache()
|
44 |
+
|
45 |
+
PandasDataFrame = TypeVar('pd.core.frame.DataFrame')
|
46 |
+
|
47 |
+
embeddings = None # global variable setup
|
48 |
+
vectorstore = None # global variable setup
|
49 |
+
model_type = None # global variable setup
|
50 |
+
|
51 |
+
max_memory_length = 0 # How long should the memory of the conversation last?
|
52 |
+
|
53 |
+
full_text = "" # Define dummy source text (full text) just to enable highlight function to load
|
54 |
+
|
55 |
+
model = [] # Define empty list for model functions to run
|
56 |
+
tokenizer = [] # Define empty list for model functions to run
|
57 |
+
|
58 |
+
## Highlight text constants
|
59 |
+
hlt_chunk_size = 12
|
60 |
+
hlt_strat = [" ", ". ", "! ", "? ", ": ", "\n\n", "\n", ", "]
|
61 |
+
hlt_overlap = 4
|
62 |
+
|
63 |
+
## Initialise NER model ##
|
64 |
+
ner_model = []#SpanMarkerModel.from_pretrained("tomaarsen/span-marker-mbert-base-multinerd") # Not currently used
|
65 |
+
|
66 |
+
## Initialise keyword model ##
|
67 |
+
# Used to pull out keywords from chat history to add to user queries behind the scenes
|
68 |
+
kw_model = pipeline("feature-extraction", model="sentence-transformers/all-MiniLM-L6-v2")
|
69 |
+
|
70 |
+
# Currently set gpu_layers to 0 even with cuda due to persistent bugs in implementation with cuda
|
71 |
+
if torch.cuda.is_available():
|
72 |
+
torch_device = "cuda"
|
73 |
+
gpu_layers = 0
|
74 |
+
else:
|
75 |
+
torch_device = "cpu"
|
76 |
+
gpu_layers = 0
|
77 |
+
|
78 |
+
print("Running on device:", torch_device)
|
79 |
+
threads = 8 #torch.get_num_threads()
|
80 |
+
print("CPU threads:", threads)
|
81 |
+
|
82 |
+
# Flan Alpaca (small, fast) Model parameters
|
83 |
+
temperature: float = 0.1
|
84 |
+
top_k: int = 3
|
85 |
+
top_p: float = 1
|
86 |
+
repetition_penalty: float = 1.3
|
87 |
+
flan_alpaca_repetition_penalty: float = 1.3
|
88 |
+
last_n_tokens: int = 64
|
89 |
+
max_new_tokens: int = 256
|
90 |
+
seed: int = 42
|
91 |
+
reset: bool = False
|
92 |
+
stream: bool = True
|
93 |
+
threads: int = threads
|
94 |
+
batch_size:int = 256
|
95 |
+
context_length:int = 2048
|
96 |
+
sample = True
|
97 |
+
|
98 |
+
|
99 |
+
class CtransInitConfig_gpu:
|
100 |
+
def __init__(self, temperature=temperature,
|
101 |
+
top_k=top_k,
|
102 |
+
top_p=top_p,
|
103 |
+
repetition_penalty=repetition_penalty,
|
104 |
+
last_n_tokens=last_n_tokens,
|
105 |
+
max_new_tokens=max_new_tokens,
|
106 |
+
seed=seed,
|
107 |
+
reset=reset,
|
108 |
+
stream=stream,
|
109 |
+
threads=threads,
|
110 |
+
batch_size=batch_size,
|
111 |
+
context_length=context_length,
|
112 |
+
gpu_layers=gpu_layers):
|
113 |
+
self.temperature = temperature
|
114 |
+
self.top_k = top_k
|
115 |
+
self.top_p = top_p
|
116 |
+
self.repetition_penalty = repetition_penalty# repetition_penalty
|
117 |
+
self.last_n_tokens = last_n_tokens
|
118 |
+
self.max_new_tokens = max_new_tokens
|
119 |
+
self.seed = seed
|
120 |
+
self.reset = reset
|
121 |
+
self.stream = stream
|
122 |
+
self.threads = threads
|
123 |
+
self.batch_size = batch_size
|
124 |
+
self.context_length = context_length
|
125 |
+
self.gpu_layers = gpu_layers
|
126 |
+
# self.stop: list[str] = field(default_factory=lambda: [stop_string])
|
127 |
+
|
128 |
+
def update_gpu(self, new_value):
|
129 |
+
self.gpu_layers = new_value
|
130 |
+
|
131 |
+
class CtransInitConfig_cpu(CtransInitConfig_gpu):
|
132 |
+
def __init__(self):
|
133 |
+
super().__init__()
|
134 |
+
self.gpu_layers = 0
|
135 |
+
|
136 |
+
gpu_config = CtransInitConfig_gpu()
|
137 |
+
cpu_config = CtransInitConfig_cpu()
|
138 |
+
|
139 |
+
|
140 |
+
class CtransGenGenerationConfig:
|
141 |
+
def __init__(self, temperature=temperature,
|
142 |
+
top_k=top_k,
|
143 |
+
top_p=top_p,
|
144 |
+
repetition_penalty=repetition_penalty,
|
145 |
+
last_n_tokens=last_n_tokens,
|
146 |
+
seed=seed,
|
147 |
+
threads=threads,
|
148 |
+
batch_size=batch_size,
|
149 |
+
reset=True
|
150 |
+
):
|
151 |
+
self.temperature = temperature
|
152 |
+
self.top_k = top_k
|
153 |
+
self.top_p = top_p
|
154 |
+
self.repetition_penalty = repetition_penalty# repetition_penalty
|
155 |
+
self.last_n_tokens = last_n_tokens
|
156 |
+
self.seed = seed
|
157 |
+
self.threads = threads
|
158 |
+
self.batch_size = batch_size
|
159 |
+
self.reset = reset
|
160 |
+
|
161 |
+
def update_temp(self, new_value):
|
162 |
+
self.temperature = new_value
|
163 |
+
|
164 |
+
# Vectorstore funcs
|
165 |
+
|
166 |
+
def docs_to_faiss_save(docs_out:PandasDataFrame, embeddings=embeddings):
|
167 |
+
|
168 |
+
print(f"> Total split documents: {len(docs_out)}")
|
169 |
+
|
170 |
+
vectorstore_func = FAISS.from_documents(documents=docs_out, embedding=embeddings)
|
171 |
+
|
172 |
+
'''
|
173 |
+
#with open("vectorstore.pkl", "wb") as f:
|
174 |
+
#pickle.dump(vectorstore, f)
|
175 |
+
'''
|
176 |
+
|
177 |
+
#if Path(save_to).exists():
|
178 |
+
# vectorstore_func.save_local(folder_path=save_to)
|
179 |
+
#else:
|
180 |
+
# os.mkdir(save_to)
|
181 |
+
# vectorstore_func.save_local(folder_path=save_to)
|
182 |
+
|
183 |
+
global vectorstore
|
184 |
+
|
185 |
+
vectorstore = vectorstore_func
|
186 |
+
|
187 |
+
out_message = "Document processing complete"
|
188 |
+
|
189 |
+
#print(out_message)
|
190 |
+
#print(f"> Saved to: {save_to}")
|
191 |
+
|
192 |
+
return out_message
|
193 |
+
|
194 |
+
# Prompt functions
|
195 |
+
|
196 |
+
def base_prompt_templates(model_type = "Flan Alpaca (small, fast)"):
|
197 |
+
|
198 |
+
#EXAMPLE_PROMPT = PromptTemplate(
|
199 |
+
# template="\nCONTENT:\n\n{page_content}\n\nSOURCE: {source}\n\n",
|
200 |
+
# input_variables=["page_content", "source"],
|
201 |
+
#)
|
202 |
+
|
203 |
+
CONTENT_PROMPT = PromptTemplate(
|
204 |
+
template="{page_content}\n\n",#\n\nSOURCE: {source}\n\n",
|
205 |
+
input_variables=["page_content"]
|
206 |
+
)
|
207 |
+
|
208 |
+
# The main prompt:
|
209 |
+
|
210 |
+
instruction_prompt_template_alpaca_quote = """### Instruction:
|
211 |
+
Quote directly from the SOURCE below that best answers the QUESTION. Only quote full sentences in the correct order. If you cannot find an answer, start your response with "My best guess is: ".
|
212 |
+
|
213 |
+
CONTENT: {summaries}
|
214 |
+
QUESTION: {question}
|
215 |
+
|
216 |
+
Response:"""
|
217 |
+
|
218 |
+
instruction_prompt_template_alpaca = """### Instruction:
|
219 |
+
### User:
|
220 |
+
Answer the QUESTION using information from the following CONTENT.
|
221 |
+
CONTENT: {summaries}
|
222 |
+
QUESTION: {question}
|
223 |
+
|
224 |
+
Response:"""
|
225 |
+
|
226 |
+
|
227 |
+
instruction_prompt_template_wizard_orca = """### HUMAN:
|
228 |
+
Answer the QUESTION below based on the CONTENT. Only refer to CONTENT that directly answers the question.
|
229 |
+
CONTENT - {summaries}
|
230 |
+
QUESTION - {question}
|
231 |
+
### RESPONSE:
|
232 |
+
"""
|
233 |
+
|
234 |
+
|
235 |
+
instruction_prompt_template_orca = """
|
236 |
+
### System:
|
237 |
+
You are an AI assistant that follows instruction extremely well. Help as much as you can.
|
238 |
+
### User:
|
239 |
+
Answer the QUESTION with a short response using information from the following CONTENT.
|
240 |
+
QUESTION: {question}
|
241 |
+
CONTENT: {summaries}
|
242 |
+
|
243 |
+
### Response:"""
|
244 |
+
|
245 |
+
instruction_prompt_template_orca_quote = """
|
246 |
+
### System:
|
247 |
+
You are an AI assistant that follows instruction extremely well. Help as much as you can.
|
248 |
+
### User:
|
249 |
+
Quote text from the CONTENT to answer the QUESTION below.
|
250 |
+
QUESTION: {question}
|
251 |
+
CONTENT: {summaries}
|
252 |
+
### Response:
|
253 |
+
"""
|
254 |
+
|
255 |
+
|
256 |
+
instruction_prompt_mistral_orca = """<|im_start|>system\n
|
257 |
+
You are an AI assistant that follows instruction extremely well. Help as much as you can.
|
258 |
+
<|im_start|>user\n
|
259 |
+
Answer the QUESTION using information from the following CONTENT. Respond with short answers that directly answer the question.
|
260 |
+
CONTENT: {summaries}
|
261 |
+
QUESTION: {question}\n
|
262 |
+
Answer:<|im_end|>"""
|
263 |
+
|
264 |
+
if model_type == "Flan Alpaca (small, fast)":
|
265 |
+
INSTRUCTION_PROMPT=PromptTemplate(template=instruction_prompt_template_alpaca, input_variables=['question', 'summaries'])
|
266 |
+
elif model_type == "Mistral Open Orca (larger, slow)":
|
267 |
+
INSTRUCTION_PROMPT=PromptTemplate(template=instruction_prompt_mistral_orca, input_variables=['question', 'summaries'])
|
268 |
+
|
269 |
+
return INSTRUCTION_PROMPT, CONTENT_PROMPT
|
270 |
+
|
271 |
+
def write_out_metadata_as_string(metadata_in):
|
272 |
+
metadata_string = [f"{' '.join(f'{k}: {v}' for k, v in d.items() if k != 'page_section')}" for d in metadata_in] # ['metadata']
|
273 |
+
return metadata_string
|
274 |
+
|
275 |
+
def generate_expanded_prompt(inputs: Dict[str, str], instruction_prompt, content_prompt, extracted_memory, vectorstore, embeddings, out_passages = 2): # ,
|
276 |
+
|
277 |
+
question = inputs["question"]
|
278 |
+
chat_history = inputs["chat_history"]
|
279 |
+
|
280 |
+
|
281 |
+
new_question_kworded = adapt_q_from_chat_history(question, chat_history, extracted_memory) # new_question_keywords,
|
282 |
+
|
283 |
+
|
284 |
+
docs_keep_as_doc, doc_df, docs_keep_out = hybrid_retrieval(new_question_kworded, vectorstore, embeddings, k_val = 25, out_passages = out_passages,
|
285 |
+
vec_score_cut_off = 0.85, vec_weight = 1, bm25_weight = 1, svm_weight = 1)#,
|
286 |
+
#vectorstore=globals()["vectorstore"], embeddings=globals()["embeddings"])
|
287 |
+
|
288 |
+
#print(docs_keep_as_doc)
|
289 |
+
#print(doc_df)
|
290 |
+
if (not docs_keep_as_doc) | (doc_df.empty):
|
291 |
+
sorry_prompt = """Say 'Sorry, there is no relevant information to answer this question.'.
|
292 |
+
RESPONSE:"""
|
293 |
+
return sorry_prompt, "No relevant sources found.", new_question_kworded
|
294 |
+
|
295 |
+
# Expand the found passages to the neighbouring context
|
296 |
+
file_type = determine_file_type(doc_df['meta_url'][0])
|
297 |
+
|
298 |
+
# Only expand passages if not tabular data
|
299 |
+
if (file_type != ".csv") & (file_type != ".xlsx"):
|
300 |
+
docs_keep_as_doc, doc_df = get_expanded_passages(vectorstore, docs_keep_out, width=3)
|
301 |
+
|
302 |
+
|
303 |
+
|
304 |
+
# Build up sources content to add to user display
|
305 |
+
doc_df['meta_clean'] = write_out_metadata_as_string(doc_df["metadata"]) # [f"<b>{' '.join(f'{k}: {v}' for k, v in d.items() if k != 'page_section')}</b>" for d in doc_df['metadata']]
|
306 |
+
|
307 |
+
# Remove meta text from the page content if it already exists there
|
308 |
+
doc_df['page_content_no_meta'] = doc_df.apply(lambda row: row['page_content'].replace(row['meta_clean'] + ". ", ""), axis=1)
|
309 |
+
doc_df['content_meta'] = doc_df['meta_clean'].astype(str) + ".<br><br>" + doc_df['page_content_no_meta'].astype(str)
|
310 |
+
|
311 |
+
#modified_page_content = [f" Document {i+1} - {word}" for i, word in enumerate(doc_df['page_content'])]
|
312 |
+
modified_page_content = [f" Document {i+1} - {word}" for i, word in enumerate(doc_df['content_meta'])]
|
313 |
+
docs_content_string = '<br><br>'.join(modified_page_content)
|
314 |
+
|
315 |
+
sources_docs_content_string = '<br><br>'.join(doc_df['content_meta'])#.replace(" "," ")#.strip()
|
316 |
+
|
317 |
+
instruction_prompt_out = instruction_prompt.format(question=new_question_kworded, summaries=docs_content_string)
|
318 |
+
|
319 |
+
print('Final prompt is: ')
|
320 |
+
print(instruction_prompt_out)
|
321 |
+
|
322 |
+
return instruction_prompt_out, sources_docs_content_string, new_question_kworded
|
323 |
+
|
324 |
+
def create_full_prompt(user_input, history, extracted_memory, vectorstore, embeddings, model_type, out_passages):
|
325 |
+
|
326 |
+
if not user_input.strip():
|
327 |
+
return history, "", "Respond with 'Please enter a question.' RESPONSE:"
|
328 |
+
|
329 |
+
#if chain_agent is None:
|
330 |
+
# history.append((user_input, "Please click the button to submit the Huggingface API key before using the chatbot (top right)"))
|
331 |
+
# return history, history, "", ""
|
332 |
+
print("\n==== date/time: " + str(datetime.datetime.now()) + " ====")
|
333 |
+
print("User input: " + user_input)
|
334 |
+
|
335 |
+
history = history or []
|
336 |
+
|
337 |
+
# Create instruction prompt
|
338 |
+
instruction_prompt, content_prompt = base_prompt_templates(model_type=model_type)
|
339 |
+
instruction_prompt_out, docs_content_string, new_question_kworded =\
|
340 |
+
generate_expanded_prompt({"question": user_input, "chat_history": history}, #vectorstore,
|
341 |
+
instruction_prompt, content_prompt, extracted_memory, vectorstore, embeddings, out_passages)
|
342 |
+
|
343 |
+
|
344 |
+
history.append(user_input)
|
345 |
+
|
346 |
+
print("Output history is:")
|
347 |
+
print(history)
|
348 |
+
|
349 |
+
print("Final prompt to model is:")
|
350 |
+
print(instruction_prompt_out)
|
351 |
+
|
352 |
+
return history, docs_content_string, instruction_prompt_out
|
353 |
+
|
354 |
+
# Chat functions
|
355 |
+
def produce_streaming_answer_chatbot(history, full_prompt, model_type,
|
356 |
+
temperature=temperature,
|
357 |
+
max_new_tokens=max_new_tokens,
|
358 |
+
sample=sample,
|
359 |
+
repetition_penalty=repetition_penalty,
|
360 |
+
top_p=top_p,
|
361 |
+
top_k=top_k
|
362 |
+
):
|
363 |
+
#print("Model type is: ", model_type)
|
364 |
+
|
365 |
+
#if not full_prompt.strip():
|
366 |
+
# if history is None:
|
367 |
+
# history = []
|
368 |
+
|
369 |
+
# return history
|
370 |
+
|
371 |
+
if model_type == "Flan Alpaca (small, fast)":
|
372 |
+
# Get the model and tokenizer, and tokenize the user text.
|
373 |
+
model_inputs = tokenizer(text=full_prompt, return_tensors="pt", return_attention_mask=False).to(torch_device) # return_attention_mask=False was added
|
374 |
+
|
375 |
+
# Start generation on a separate thread, so that we don't block the UI. The text is pulled from the streamer
|
376 |
+
# in the main thread. Adds timeout to the streamer to handle exceptions in the generation thread.
|
377 |
+
streamer = TextIteratorStreamer(tokenizer, timeout=120., skip_prompt=True, skip_special_tokens=True)
|
378 |
+
generate_kwargs = dict(
|
379 |
+
model_inputs,
|
380 |
+
streamer=streamer,
|
381 |
+
max_new_tokens=max_new_tokens,
|
382 |
+
do_sample=sample,
|
383 |
+
repetition_penalty=repetition_penalty,
|
384 |
+
top_p=top_p,
|
385 |
+
temperature=temperature,
|
386 |
+
top_k=top_k
|
387 |
+
)
|
388 |
+
|
389 |
+
print(generate_kwargs)
|
390 |
+
|
391 |
+
t = Thread(target=model.generate, kwargs=generate_kwargs)
|
392 |
+
t.start()
|
393 |
+
|
394 |
+
# Pull the generated text from the streamer, and update the model output.
|
395 |
+
start = time.time()
|
396 |
+
NUM_TOKENS=0
|
397 |
+
print('-'*4+'Start Generation'+'-'*4)
|
398 |
+
|
399 |
+
history[-1][1] = ""
|
400 |
+
for new_text in streamer:
|
401 |
+
if new_text == None: new_text = ""
|
402 |
+
history[-1][1] += new_text
|
403 |
+
NUM_TOKENS+=1
|
404 |
+
yield history
|
405 |
+
|
406 |
+
time_generate = time.time() - start
|
407 |
+
print('\n')
|
408 |
+
print('-'*4+'End Generation'+'-'*4)
|
409 |
+
print(f'Num of generated tokens: {NUM_TOKENS}')
|
410 |
+
print(f'Time for complete generation: {time_generate}s')
|
411 |
+
print(f'Tokens per secound: {NUM_TOKENS/time_generate}')
|
412 |
+
print(f'Time per token: {(time_generate/NUM_TOKENS)*1000}ms')
|
413 |
+
|
414 |
+
elif model_type == "Mistral Open Orca (larger, slow)":
|
415 |
+
tokens = model.tokenize(full_prompt)
|
416 |
+
|
417 |
+
gen_config = CtransGenGenerationConfig()
|
418 |
+
gen_config.update_temp(temperature)
|
419 |
+
|
420 |
+
print(vars(gen_config))
|
421 |
+
|
422 |
+
# Pull the generated text from the streamer, and update the model output.
|
423 |
+
start = time.time()
|
424 |
+
NUM_TOKENS=0
|
425 |
+
print('-'*4+'Start Generation'+'-'*4)
|
426 |
+
|
427 |
+
history[-1][1] = ""
|
428 |
+
for new_text in model.generate(tokens, **vars(gen_config)): #CtransGen_generate(prompt=full_prompt)#, config=CtransGenGenerationConfig()): # #top_k=top_k, temperature=temperature, repetition_penalty=repetition_penalty,
|
429 |
+
if new_text == None: new_text = ""
|
430 |
+
history[-1][1] += model.detokenize(new_text) #new_text
|
431 |
+
NUM_TOKENS+=1
|
432 |
+
yield history
|
433 |
+
|
434 |
+
time_generate = time.time() - start
|
435 |
+
print('\n')
|
436 |
+
print('-'*4+'End Generation'+'-'*4)
|
437 |
+
print(f'Num of generated tokens: {NUM_TOKENS}')
|
438 |
+
print(f'Time for complete generation: {time_generate}s')
|
439 |
+
print(f'Tokens per secound: {NUM_TOKENS/time_generate}')
|
440 |
+
print(f'Time per token: {(time_generate/NUM_TOKENS)*1000}ms')
|
441 |
+
|
442 |
+
# Chat helper functions
|
443 |
+
|
444 |
+
def adapt_q_from_chat_history(question, chat_history, extracted_memory, keyword_model=""):#keyword_model): # new_question_keywords,
|
445 |
+
|
446 |
+
chat_history_str, chat_history_first_q, chat_history_first_ans, max_memory_length = _get_chat_history(chat_history)
|
447 |
+
|
448 |
+
if chat_history_str:
|
449 |
+
# Keyword extraction is now done in the add_inputs_to_history function
|
450 |
+
#remove_q_stopwords(str(chat_history_first_q) + " " + str(chat_history_first_ans))
|
451 |
+
|
452 |
+
|
453 |
+
new_question_kworded = str(extracted_memory) + ". " + question #+ " " + new_question_keywords
|
454 |
+
#extracted_memory + " " + question
|
455 |
+
|
456 |
+
else:
|
457 |
+
new_question_kworded = question #new_question_keywords
|
458 |
+
|
459 |
+
#print("Question output is: " + new_question_kworded)
|
460 |
+
|
461 |
+
return new_question_kworded
|
462 |
+
|
463 |
+
def determine_file_type(file_path):
|
464 |
+
"""
|
465 |
+
Determine the file type based on its extension.
|
466 |
+
|
467 |
+
Parameters:
|
468 |
+
file_path (str): Path to the file.
|
469 |
+
|
470 |
+
Returns:
|
471 |
+
str: File extension (e.g., '.pdf', '.docx', '.txt', '.html').
|
472 |
+
"""
|
473 |
+
return os.path.splitext(file_path)[1].lower()
|
474 |
+
|
475 |
+
|
476 |
+
def create_doc_df(docs_keep_out):
|
477 |
+
# Extract content and metadata from 'winning' passages.
|
478 |
+
content=[]
|
479 |
+
meta=[]
|
480 |
+
meta_url=[]
|
481 |
+
page_section=[]
|
482 |
+
score=[]
|
483 |
+
|
484 |
+
doc_df = pd.DataFrame()
|
485 |
+
|
486 |
+
|
487 |
+
|
488 |
+
for item in docs_keep_out:
|
489 |
+
content.append(item[0].page_content)
|
490 |
+
meta.append(item[0].metadata)
|
491 |
+
meta_url.append(item[0].metadata['source'])
|
492 |
+
|
493 |
+
file_extension = determine_file_type(item[0].metadata['source'])
|
494 |
+
if (file_extension != ".csv") & (file_extension != ".xlsx"):
|
495 |
+
page_section.append(item[0].metadata['page_section'])
|
496 |
+
else: page_section.append("")
|
497 |
+
score.append(item[1])
|
498 |
+
|
499 |
+
# Create df from 'winning' passages
|
500 |
+
|
501 |
+
doc_df = pd.DataFrame(list(zip(content, meta, page_section, meta_url, score)),
|
502 |
+
columns =['page_content', 'metadata', 'page_section', 'meta_url', 'score'])
|
503 |
+
|
504 |
+
docs_content = doc_df['page_content'].astype(str)
|
505 |
+
doc_df['full_url'] = "https://" + doc_df['meta_url']
|
506 |
+
|
507 |
+
return doc_df
|
508 |
+
|
509 |
+
def hybrid_retrieval(new_question_kworded, vectorstore, embeddings, k_val, out_passages,
|
510 |
+
vec_score_cut_off, vec_weight, bm25_weight, svm_weight): # ,vectorstore, embeddings
|
511 |
+
|
512 |
+
#vectorstore=globals()["vectorstore"]
|
513 |
+
#embeddings=globals()["embeddings"]
|
514 |
+
doc_df = pd.DataFrame()
|
515 |
+
|
516 |
+
|
517 |
+
docs = vectorstore.similarity_search_with_score(new_question_kworded, k=k_val)
|
518 |
+
|
519 |
+
print("Docs from similarity search:")
|
520 |
+
print(docs)
|
521 |
+
|
522 |
+
# Keep only documents with a certain score
|
523 |
+
docs_len = [len(x[0].page_content) for x in docs]
|
524 |
+
docs_scores = [x[1] for x in docs]
|
525 |
+
|
526 |
+
# Only keep sources that are sufficiently relevant (i.e. similarity search score below threshold below)
|
527 |
+
score_more_limit = pd.Series(docs_scores) < vec_score_cut_off
|
528 |
+
docs_keep = list(compress(docs, score_more_limit))
|
529 |
+
|
530 |
+
if not docs_keep:
|
531 |
+
return [], pd.DataFrame(), []
|
532 |
+
|
533 |
+
# Only keep sources that are at least 100 characters long
|
534 |
+
length_more_limit = pd.Series(docs_len) >= 100
|
535 |
+
docs_keep = list(compress(docs_keep, length_more_limit))
|
536 |
+
|
537 |
+
if not docs_keep:
|
538 |
+
return [], pd.DataFrame(), []
|
539 |
+
|
540 |
+
docs_keep_as_doc = [x[0] for x in docs_keep]
|
541 |
+
docs_keep_length = len(docs_keep_as_doc)
|
542 |
+
|
543 |
+
|
544 |
+
|
545 |
+
if docs_keep_length == 1:
|
546 |
+
|
547 |
+
content=[]
|
548 |
+
meta_url=[]
|
549 |
+
score=[]
|
550 |
+
|
551 |
+
for item in docs_keep:
|
552 |
+
content.append(item[0].page_content)
|
553 |
+
meta_url.append(item[0].metadata['source'])
|
554 |
+
score.append(item[1])
|
555 |
+
|
556 |
+
# Create df from 'winning' passages
|
557 |
+
|
558 |
+
doc_df = pd.DataFrame(list(zip(content, meta_url, score)),
|
559 |
+
columns =['page_content', 'meta_url', 'score'])
|
560 |
+
|
561 |
+
docs_content = doc_df['page_content'].astype(str)
|
562 |
+
docs_url = doc_df['meta_url']
|
563 |
+
|
564 |
+
return docs_keep_as_doc, docs_content, docs_url
|
565 |
+
|
566 |
+
# Check for if more docs are removed than the desired output
|
567 |
+
if out_passages > docs_keep_length:
|
568 |
+
out_passages = docs_keep_length
|
569 |
+
k_val = docs_keep_length
|
570 |
+
|
571 |
+
vec_rank = [*range(1, docs_keep_length+1)]
|
572 |
+
vec_score = [(docs_keep_length/x)*vec_weight for x in vec_rank]
|
573 |
+
|
574 |
+
# 2nd level check on retrieved docs with BM25
|
575 |
+
|
576 |
+
content_keep=[]
|
577 |
+
for item in docs_keep:
|
578 |
+
content_keep.append(item[0].page_content)
|
579 |
+
|
580 |
+
corpus = corpus = [doc.lower().split() for doc in content_keep]
|
581 |
+
dictionary = Dictionary(corpus)
|
582 |
+
bm25_model = OkapiBM25Model(dictionary=dictionary)
|
583 |
+
bm25_corpus = bm25_model[list(map(dictionary.doc2bow, corpus))]
|
584 |
+
bm25_index = SparseMatrixSimilarity(bm25_corpus, num_docs=len(corpus), num_terms=len(dictionary),
|
585 |
+
normalize_queries=False, normalize_documents=False)
|
586 |
+
query = new_question_kworded.lower().split()
|
587 |
+
tfidf_model = TfidfModel(dictionary=dictionary, smartirs='bnn') # Enforce binary weighting of queries
|
588 |
+
tfidf_query = tfidf_model[dictionary.doc2bow(query)]
|
589 |
+
similarities = np.array(bm25_index[tfidf_query])
|
590 |
+
#print(similarities)
|
591 |
+
temp = similarities.argsort()
|
592 |
+
ranks = np.arange(len(similarities))[temp.argsort()][::-1]
|
593 |
+
|
594 |
+
# Pair each index with its corresponding value
|
595 |
+
pairs = list(zip(ranks, docs_keep_as_doc))
|
596 |
+
# Sort the pairs by the indices
|
597 |
+
pairs.sort()
|
598 |
+
# Extract the values in the new order
|
599 |
+
bm25_result = [value for ranks, value in pairs]
|
600 |
+
|
601 |
+
bm25_rank=[]
|
602 |
+
bm25_score = []
|
603 |
+
|
604 |
+
for vec_item in docs_keep:
|
605 |
+
x = 0
|
606 |
+
for bm25_item in bm25_result:
|
607 |
+
x = x + 1
|
608 |
+
if bm25_item.page_content == vec_item[0].page_content:
|
609 |
+
bm25_rank.append(x)
|
610 |
+
bm25_score.append((docs_keep_length/x)*bm25_weight)
|
611 |
+
|
612 |
+
# 3rd level check on retrieved docs with SVM retriever
|
613 |
+
svm_retriever = SVMRetriever.from_texts(content_keep, embeddings, k = k_val)
|
614 |
+
svm_result = svm_retriever.get_relevant_documents(new_question_kworded)
|
615 |
+
|
616 |
+
|
617 |
+
svm_rank=[]
|
618 |
+
svm_score = []
|
619 |
+
|
620 |
+
for vec_item in docs_keep:
|
621 |
+
x = 0
|
622 |
+
for svm_item in svm_result:
|
623 |
+
x = x + 1
|
624 |
+
if svm_item.page_content == vec_item[0].page_content:
|
625 |
+
svm_rank.append(x)
|
626 |
+
svm_score.append((docs_keep_length/x)*svm_weight)
|
627 |
+
|
628 |
+
|
629 |
+
## Calculate final score based on three ranking methods
|
630 |
+
final_score = [a + b + c for a, b, c in zip(vec_score, bm25_score, svm_score)]
|
631 |
+
final_rank = [sorted(final_score, reverse=True).index(x)+1 for x in final_score]
|
632 |
+
# Force final_rank to increment by 1 each time
|
633 |
+
final_rank = list(pd.Series(final_rank).rank(method='first'))
|
634 |
+
|
635 |
+
#print("final rank: " + str(final_rank))
|
636 |
+
#print("out_passages: " + str(out_passages))
|
637 |
+
|
638 |
+
best_rank_index_pos = []
|
639 |
+
|
640 |
+
for x in range(1,out_passages+1):
|
641 |
+
try:
|
642 |
+
best_rank_index_pos.append(final_rank.index(x))
|
643 |
+
except IndexError: # catch the error
|
644 |
+
pass
|
645 |
+
|
646 |
+
# Adjust best_rank_index_pos to
|
647 |
+
|
648 |
+
best_rank_pos_series = pd.Series(best_rank_index_pos)
|
649 |
+
|
650 |
+
|
651 |
+
docs_keep_out = [docs_keep[i] for i in best_rank_index_pos]
|
652 |
+
|
653 |
+
# Keep only 'best' options
|
654 |
+
docs_keep_as_doc = [x[0] for x in docs_keep_out]
|
655 |
+
|
656 |
+
# Make df of best options
|
657 |
+
doc_df = create_doc_df(docs_keep_out)
|
658 |
+
|
659 |
+
return docs_keep_as_doc, doc_df, docs_keep_out
|
660 |
+
|
661 |
+
def get_expanded_passages(vectorstore, docs, width):
|
662 |
+
|
663 |
+
"""
|
664 |
+
Extracts expanded passages based on given documents and a width for context.
|
665 |
+
|
666 |
+
Parameters:
|
667 |
+
- vectorstore: The primary data source.
|
668 |
+
- docs: List of documents to be expanded.
|
669 |
+
- width: Number of documents to expand around a given document for context.
|
670 |
+
|
671 |
+
Returns:
|
672 |
+
- expanded_docs: List of expanded Document objects.
|
673 |
+
- doc_df: DataFrame representation of expanded_docs.
|
674 |
+
"""
|
675 |
+
|
676 |
+
from collections import defaultdict
|
677 |
+
|
678 |
+
def get_docs_from_vstore(vectorstore):
|
679 |
+
vector = vectorstore.docstore._dict
|
680 |
+
return list(vector.items())
|
681 |
+
|
682 |
+
def extract_details(docs_list):
|
683 |
+
docs_list_out = [tup[1] for tup in docs_list]
|
684 |
+
content = [doc.page_content for doc in docs_list_out]
|
685 |
+
meta = [doc.metadata for doc in docs_list_out]
|
686 |
+
return ''.join(content), meta[0], meta[-1]
|
687 |
+
|
688 |
+
def get_parent_content_and_meta(vstore_docs, width, target):
|
689 |
+
#target_range = range(max(0, target - width), min(len(vstore_docs), target + width + 1))
|
690 |
+
target_range = range(max(0, target), min(len(vstore_docs), target + width + 1)) # Now only selects extra passages AFTER the found passage
|
691 |
+
parent_vstore_out = [vstore_docs[i] for i in target_range]
|
692 |
+
|
693 |
+
content_str_out, meta_first_out, meta_last_out = [], [], []
|
694 |
+
for _ in parent_vstore_out:
|
695 |
+
content_str, meta_first, meta_last = extract_details(parent_vstore_out)
|
696 |
+
content_str_out.append(content_str)
|
697 |
+
meta_first_out.append(meta_first)
|
698 |
+
meta_last_out.append(meta_last)
|
699 |
+
return content_str_out, meta_first_out, meta_last_out
|
700 |
+
|
701 |
+
def merge_dicts_except_source(d1, d2):
|
702 |
+
merged = {}
|
703 |
+
for key in d1:
|
704 |
+
if key != "source":
|
705 |
+
merged[key] = str(d1[key]) + " to " + str(d2[key])
|
706 |
+
else:
|
707 |
+
merged[key] = d1[key] # or d2[key], based on preference
|
708 |
+
return merged
|
709 |
+
|
710 |
+
def merge_two_lists_of_dicts(list1, list2):
|
711 |
+
return [merge_dicts_except_source(d1, d2) for d1, d2 in zip(list1, list2)]
|
712 |
+
|
713 |
+
# Step 1: Filter vstore_docs
|
714 |
+
vstore_docs = get_docs_from_vstore(vectorstore)
|
715 |
+
doc_sources = {doc.metadata['source'] for doc, _ in docs}
|
716 |
+
vstore_docs = [(k, v) for k, v in vstore_docs if v.metadata.get('source') in doc_sources]
|
717 |
+
|
718 |
+
# Step 2: Group by source and proceed
|
719 |
+
vstore_by_source = defaultdict(list)
|
720 |
+
for k, v in vstore_docs:
|
721 |
+
vstore_by_source[v.metadata['source']].append((k, v))
|
722 |
+
|
723 |
+
expanded_docs = []
|
724 |
+
for doc, score in docs:
|
725 |
+
search_source = doc.metadata['source']
|
726 |
+
|
727 |
+
|
728 |
+
#if file_type == ".csv" | file_type == ".xlsx":
|
729 |
+
# content_str, meta_first, meta_last = get_parent_content_and_meta(vstore_by_source[search_source], 0, search_index)
|
730 |
+
|
731 |
+
#else:
|
732 |
+
search_section = doc.metadata['page_section']
|
733 |
+
parent_vstore_meta_section = [doc.metadata['page_section'] for _, doc in vstore_by_source[search_source]]
|
734 |
+
search_index = parent_vstore_meta_section.index(search_section) if search_section in parent_vstore_meta_section else -1
|
735 |
+
|
736 |
+
content_str, meta_first, meta_last = get_parent_content_and_meta(vstore_by_source[search_source], width, search_index)
|
737 |
+
meta_full = merge_two_lists_of_dicts(meta_first, meta_last)
|
738 |
+
|
739 |
+
expanded_doc = (Document(page_content=content_str[0], metadata=meta_full[0]), score)
|
740 |
+
expanded_docs.append(expanded_doc)
|
741 |
+
|
742 |
+
doc_df = pd.DataFrame()
|
743 |
+
|
744 |
+
doc_df = create_doc_df(expanded_docs) # Assuming you've defined the 'create_doc_df' function elsewhere
|
745 |
+
|
746 |
+
return expanded_docs, doc_df
|
747 |
+
|
748 |
+
def highlight_found_text(search_text: str, full_text: str, hlt_chunk_size:int=hlt_chunk_size, hlt_strat:List=hlt_strat, hlt_overlap:int=hlt_overlap) -> str:
|
749 |
+
"""
|
750 |
+
Highlights occurrences of search_text within full_text.
|
751 |
+
|
752 |
+
Parameters:
|
753 |
+
- search_text (str): The text to be searched for within full_text.
|
754 |
+
- full_text (str): The text within which search_text occurrences will be highlighted.
|
755 |
+
|
756 |
+
Returns:
|
757 |
+
- str: A string with occurrences of search_text highlighted.
|
758 |
+
|
759 |
+
Example:
|
760 |
+
>>> highlight_found_text("world", "Hello, world! This is a test. Another world awaits.")
|
761 |
+
'Hello, <mark style="color:black;">world</mark>! This is a test. Another <mark style="color:black;">world</mark> awaits.'
|
762 |
+
"""
|
763 |
+
|
764 |
+
def extract_text_from_input(text, i=0):
|
765 |
+
if isinstance(text, str):
|
766 |
+
return text.replace(" ", " ").strip()
|
767 |
+
elif isinstance(text, list):
|
768 |
+
return text[i][0].replace(" ", " ").strip()
|
769 |
+
else:
|
770 |
+
return ""
|
771 |
+
|
772 |
+
def extract_search_text_from_input(text):
|
773 |
+
if isinstance(text, str):
|
774 |
+
return text.replace(" ", " ").strip()
|
775 |
+
elif isinstance(text, list):
|
776 |
+
return text[-1][1].replace(" ", " ").strip()
|
777 |
+
else:
|
778 |
+
return ""
|
779 |
+
|
780 |
+
full_text = extract_text_from_input(full_text)
|
781 |
+
search_text = extract_search_text_from_input(search_text)
|
782 |
+
|
783 |
+
|
784 |
+
|
785 |
+
text_splitter = RecursiveCharacterTextSplitter(
|
786 |
+
chunk_size=hlt_chunk_size,
|
787 |
+
separators=hlt_strat,
|
788 |
+
chunk_overlap=hlt_overlap,
|
789 |
+
)
|
790 |
+
sections = text_splitter.split_text(search_text)
|
791 |
+
|
792 |
+
found_positions = {}
|
793 |
+
for x in sections:
|
794 |
+
text_start_pos = 0
|
795 |
+
while text_start_pos != -1:
|
796 |
+
text_start_pos = full_text.find(x, text_start_pos)
|
797 |
+
if text_start_pos != -1:
|
798 |
+
found_positions[text_start_pos] = text_start_pos + len(x)
|
799 |
+
text_start_pos += 1
|
800 |
+
|
801 |
+
# Combine overlapping or adjacent positions
|
802 |
+
sorted_starts = sorted(found_positions.keys())
|
803 |
+
combined_positions = []
|
804 |
+
if sorted_starts:
|
805 |
+
current_start, current_end = sorted_starts[0], found_positions[sorted_starts[0]]
|
806 |
+
for start in sorted_starts[1:]:
|
807 |
+
if start <= (current_end + 10):
|
808 |
+
current_end = max(current_end, found_positions[start])
|
809 |
+
else:
|
810 |
+
combined_positions.append((current_start, current_end))
|
811 |
+
current_start, current_end = start, found_positions[start]
|
812 |
+
combined_positions.append((current_start, current_end))
|
813 |
+
|
814 |
+
# Construct pos_tokens
|
815 |
+
pos_tokens = []
|
816 |
+
prev_end = 0
|
817 |
+
for start, end in combined_positions:
|
818 |
+
if end-start > 15: # Only combine if there is a significant amount of matched text. Avoids picking up single words like 'and' etc.
|
819 |
+
pos_tokens.append(full_text[prev_end:start])
|
820 |
+
pos_tokens.append('<mark style="color:black;">' + full_text[start:end] + '</mark>')
|
821 |
+
prev_end = end
|
822 |
+
pos_tokens.append(full_text[prev_end:])
|
823 |
+
|
824 |
+
return "".join(pos_tokens)
|
825 |
+
|
826 |
+
|
827 |
+
# # Chat history functions
|
828 |
+
|
829 |
+
def clear_chat(chat_history_state, sources, chat_message, current_topic):
|
830 |
+
chat_history_state = []
|
831 |
+
sources = ''
|
832 |
+
chat_message = ''
|
833 |
+
current_topic = ''
|
834 |
+
|
835 |
+
return chat_history_state, sources, chat_message, current_topic
|
836 |
+
|
837 |
+
def _get_chat_history(chat_history: List[Tuple[str, str]], max_memory_length:int = max_memory_length): # Limit to last x interactions only
|
838 |
+
|
839 |
+
if (not chat_history) | (max_memory_length == 0):
|
840 |
+
chat_history = []
|
841 |
+
|
842 |
+
if len(chat_history) > max_memory_length:
|
843 |
+
chat_history = chat_history[-max_memory_length:]
|
844 |
+
|
845 |
+
#print(chat_history)
|
846 |
+
|
847 |
+
first_q = ""
|
848 |
+
first_ans = ""
|
849 |
+
for human_s, ai_s in chat_history:
|
850 |
+
first_q = human_s
|
851 |
+
first_ans = ai_s
|
852 |
+
|
853 |
+
#print("Text to keyword extract: " + first_q + " " + first_ans)
|
854 |
+
break
|
855 |
+
|
856 |
+
conversation = ""
|
857 |
+
for human_s, ai_s in chat_history:
|
858 |
+
human = f"Human: " + human_s
|
859 |
+
ai = f"Assistant: " + ai_s
|
860 |
+
conversation += "\n" + "\n".join([human, ai])
|
861 |
+
|
862 |
+
return conversation, first_q, first_ans, max_memory_length
|
863 |
+
|
864 |
+
def add_inputs_answer_to_history(user_message, history, current_topic):
|
865 |
+
|
866 |
+
if history is None:
|
867 |
+
history = [("","")]
|
868 |
+
|
869 |
+
#history.append((user_message, [-1]))
|
870 |
+
|
871 |
+
chat_history_str, chat_history_first_q, chat_history_first_ans, max_memory_length = _get_chat_history(history)
|
872 |
+
|
873 |
+
|
874 |
+
# Only get the keywords for the first question and response, or do it every time if over 'max_memory_length' responses in the conversation
|
875 |
+
if (len(history) == 1) | (len(history) > max_memory_length):
|
876 |
+
|
877 |
+
#print("History after appending is:")
|
878 |
+
#print(history)
|
879 |
+
|
880 |
+
first_q_and_first_ans = str(chat_history_first_q) + " " + str(chat_history_first_ans)
|
881 |
+
#ner_memory = remove_q_ner_extractor(first_q_and_first_ans)
|
882 |
+
keywords = keybert_keywords(first_q_and_first_ans, n = 8, kw_model=kw_model)
|
883 |
+
#keywords.append(ner_memory)
|
884 |
+
|
885 |
+
# Remove duplicate words while preserving order
|
886 |
+
ordered_tokens = set()
|
887 |
+
result = []
|
888 |
+
for word in keywords:
|
889 |
+
if word not in ordered_tokens:
|
890 |
+
ordered_tokens.add(word)
|
891 |
+
result.append(word)
|
892 |
+
|
893 |
+
extracted_memory = ' '.join(result)
|
894 |
+
|
895 |
+
else: extracted_memory=current_topic
|
896 |
+
|
897 |
+
print("Extracted memory is:")
|
898 |
+
print(extracted_memory)
|
899 |
+
|
900 |
+
|
901 |
+
return history, extracted_memory
|
902 |
+
|
903 |
+
# Keyword functions
|
904 |
+
|
905 |
+
def remove_q_stopwords(question): # Remove stopwords from question. Not used at the moment
|
906 |
+
# Prepare keywords from question by removing stopwords
|
907 |
+
text = question.lower()
|
908 |
+
|
909 |
+
# Remove numbers
|
910 |
+
text = re.sub('[0-9]', '', text)
|
911 |
+
|
912 |
+
tokenizer = RegexpTokenizer(r'\w+')
|
913 |
+
text_tokens = tokenizer.tokenize(text)
|
914 |
+
#text_tokens = word_tokenize(text)
|
915 |
+
tokens_without_sw = [word for word in text_tokens if not word in stopwords]
|
916 |
+
|
917 |
+
# Remove duplicate words while preserving order
|
918 |
+
ordered_tokens = set()
|
919 |
+
result = []
|
920 |
+
for word in tokens_without_sw:
|
921 |
+
if word not in ordered_tokens:
|
922 |
+
ordered_tokens.add(word)
|
923 |
+
result.append(word)
|
924 |
+
|
925 |
+
|
926 |
+
|
927 |
+
new_question_keywords = ' '.join(result)
|
928 |
+
return new_question_keywords
|
929 |
+
|
930 |
+
def remove_q_ner_extractor(question):
|
931 |
+
|
932 |
+
predict_out = ner_model.predict(question)
|
933 |
+
|
934 |
+
|
935 |
+
|
936 |
+
predict_tokens = [' '.join(v for k, v in d.items() if k == 'span') for d in predict_out]
|
937 |
+
|
938 |
+
# Remove duplicate words while preserving order
|
939 |
+
ordered_tokens = set()
|
940 |
+
result = []
|
941 |
+
for word in predict_tokens:
|
942 |
+
if word not in ordered_tokens:
|
943 |
+
ordered_tokens.add(word)
|
944 |
+
result.append(word)
|
945 |
+
|
946 |
+
|
947 |
+
|
948 |
+
new_question_keywords = ' '.join(result).lower()
|
949 |
+
return new_question_keywords
|
950 |
+
|
951 |
+
def apply_lemmatize(text, wnl=WordNetLemmatizer()):
|
952 |
+
|
953 |
+
def prep_for_lemma(text):
|
954 |
+
|
955 |
+
# Remove numbers
|
956 |
+
text = re.sub('[0-9]', '', text)
|
957 |
+
print(text)
|
958 |
+
|
959 |
+
tokenizer = RegexpTokenizer(r'\w+')
|
960 |
+
text_tokens = tokenizer.tokenize(text)
|
961 |
+
#text_tokens = word_tokenize(text)
|
962 |
+
|
963 |
+
return text_tokens
|
964 |
+
|
965 |
+
tokens = prep_for_lemma(text)
|
966 |
+
|
967 |
+
def lem_word(word):
|
968 |
+
|
969 |
+
if len(word) > 3: out_word = wnl.lemmatize(word)
|
970 |
+
else: out_word = word
|
971 |
+
|
972 |
+
return out_word
|
973 |
+
|
974 |
+
return [lem_word(token) for token in tokens]
|
975 |
+
|
976 |
+
def keybert_keywords(text, n, kw_model):
|
977 |
+
tokens_lemma = apply_lemmatize(text)
|
978 |
+
lemmatised_text = ' '.join(tokens_lemma)
|
979 |
+
|
980 |
+
keywords_text = KeyBERT(model=kw_model).extract_keywords(lemmatised_text, stop_words='english', top_n=n,
|
981 |
+
keyphrase_ngram_range=(1, 1))
|
982 |
+
keywords_list = [item[0] for item in keywords_text]
|
983 |
+
|
984 |
+
return keywords_list
|
985 |
+
|
986 |
+
# Gradio functions
|
987 |
+
def turn_off_interactivity(user_message, history):
|
988 |
+
return gr.update(value="", interactive=False), history + [[user_message, None]]
|
989 |
+
|
990 |
+
def restore_interactivity():
|
991 |
+
return gr.update(interactive=True)
|
992 |
+
|
993 |
+
def update_message(dropdown_value):
|
994 |
+
return gr.Textbox.update(value=dropdown_value)
|
995 |
+
|
996 |
+
def hide_block():
|
997 |
+
return gr.Radio.update(visible=False)
|
998 |
+
|
999 |
+
# Vote function
|
1000 |
+
|
1001 |
+
def vote(data: gr.LikeData, chat_history, instruction_prompt_out, model_type):
|
1002 |
+
import os
|
1003 |
+
import pandas as pd
|
1004 |
+
|
1005 |
+
chat_history_last = str(str(chat_history[-1][0]) + " - " + str(chat_history[-1][1]))
|
1006 |
+
|
1007 |
+
response_df = pd.DataFrame(data={"thumbs_up":data.liked,
|
1008 |
+
"chosen_response":data.value,
|
1009 |
+
"input_prompt":instruction_prompt_out,
|
1010 |
+
"chat_history":chat_history_last,
|
1011 |
+
"model_type": model_type,
|
1012 |
+
"date_time": pd.Timestamp.now()}, index=[0])
|
1013 |
+
|
1014 |
+
if data.liked:
|
1015 |
+
print("You upvoted this response: " + data.value)
|
1016 |
+
|
1017 |
+
if os.path.isfile("thumbs_up_data.csv"):
|
1018 |
+
existing_thumbs_up_df = pd.read_csv("thumbs_up_data.csv")
|
1019 |
+
thumbs_up_df_concat = pd.concat([existing_thumbs_up_df, response_df], ignore_index=True).drop("Unnamed: 0",axis=1, errors="ignore")
|
1020 |
+
thumbs_up_df_concat.to_csv("thumbs_up_data.csv")
|
1021 |
+
else:
|
1022 |
+
response_df.to_csv("thumbs_up_data.csv")
|
1023 |
+
|
1024 |
+
else:
|
1025 |
+
print("You downvoted this response: " + data.value)
|
1026 |
+
|
1027 |
+
if os.path.isfile("thumbs_down_data.csv"):
|
1028 |
+
existing_thumbs_down_df = pd.read_csv("thumbs_down_data.csv")
|
1029 |
+
thumbs_down_df_concat = pd.concat([existing_thumbs_down_df, response_df], ignore_index=True).drop("Unnamed: 0",axis=1, errors="ignore")
|
1030 |
+
thumbs_down_df_concat.to_csv("thumbs_down_data.csv")
|
1031 |
+
else:
|
1032 |
+
response_df.to_csv("thumbs_down_data.csv")
|
chatfuncs/ingest.py
ADDED
@@ -0,0 +1,655 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# ---
|
2 |
+
# jupyter:
|
3 |
+
# jupytext:
|
4 |
+
# formats: ipynb,py:light
|
5 |
+
# text_representation:
|
6 |
+
# extension: .py
|
7 |
+
# format_name: light
|
8 |
+
# format_version: '1.5'
|
9 |
+
# jupytext_version: 1.14.6
|
10 |
+
# kernelspec:
|
11 |
+
# display_name: Python 3 (ipykernel)
|
12 |
+
# language: python
|
13 |
+
# name: python3
|
14 |
+
# ---
|
15 |
+
|
16 |
+
# # Ingest website to FAISS
|
17 |
+
|
18 |
+
# ## Install/ import stuff we need
|
19 |
+
|
20 |
+
import os
|
21 |
+
from pathlib import Path
|
22 |
+
import re
|
23 |
+
import requests
|
24 |
+
import pandas as pd
|
25 |
+
import dateutil.parser
|
26 |
+
from typing import TypeVar, List
|
27 |
+
|
28 |
+
from langchain.embeddings import HuggingFaceEmbeddings # HuggingFaceInstructEmbeddings,
|
29 |
+
from langchain.vectorstores.faiss import FAISS
|
30 |
+
from langchain.vectorstores import Chroma
|
31 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
32 |
+
from langchain.docstore.document import Document
|
33 |
+
|
34 |
+
from bs4 import BeautifulSoup
|
35 |
+
from docx import Document as Doc
|
36 |
+
from pypdf import PdfReader
|
37 |
+
|
38 |
+
PandasDataFrame = TypeVar('pd.core.frame.DataFrame')
|
39 |
+
# -
|
40 |
+
|
41 |
+
split_strat = ["\n\n", "\n", ". ", "! ", "? "]
|
42 |
+
chunk_size = 500
|
43 |
+
chunk_overlap = 0
|
44 |
+
start_index = True
|
45 |
+
|
46 |
+
## Parse files
|
47 |
+
def determine_file_type(file_path):
|
48 |
+
"""
|
49 |
+
Determine the file type based on its extension.
|
50 |
+
|
51 |
+
Parameters:
|
52 |
+
file_path (str): Path to the file.
|
53 |
+
|
54 |
+
Returns:
|
55 |
+
str: File extension (e.g., '.pdf', '.docx', '.txt', '.html').
|
56 |
+
"""
|
57 |
+
return os.path.splitext(file_path)[1].lower()
|
58 |
+
|
59 |
+
def parse_file(file_paths, text_column='text'):
|
60 |
+
"""
|
61 |
+
Accepts a list of file paths, determines each file's type based on its extension,
|
62 |
+
and passes it to the relevant parsing function.
|
63 |
+
|
64 |
+
Parameters:
|
65 |
+
file_paths (list): List of file paths.
|
66 |
+
text_column (str): Name of the column in CSV/Excel files that contains the text content.
|
67 |
+
|
68 |
+
Returns:
|
69 |
+
dict: A dictionary with file paths as keys and their parsed content (or error message) as values.
|
70 |
+
"""
|
71 |
+
|
72 |
+
|
73 |
+
|
74 |
+
if not isinstance(file_paths, list):
|
75 |
+
raise ValueError("Expected a list of file paths.")
|
76 |
+
|
77 |
+
extension_to_parser = {
|
78 |
+
'.pdf': parse_pdf,
|
79 |
+
'.docx': parse_docx,
|
80 |
+
'.txt': parse_txt,
|
81 |
+
'.html': parse_html,
|
82 |
+
'.htm': parse_html, # Considering both .html and .htm for HTML files
|
83 |
+
'.csv': lambda file_path: parse_csv_or_excel(file_path, text_column),
|
84 |
+
'.xlsx': lambda file_path: parse_csv_or_excel(file_path, text_column)
|
85 |
+
}
|
86 |
+
|
87 |
+
parsed_contents = {}
|
88 |
+
file_names = []
|
89 |
+
|
90 |
+
for file_path in file_paths:
|
91 |
+
print(file_path.name)
|
92 |
+
#file = open(file_path.name, 'r')
|
93 |
+
#print(file)
|
94 |
+
file_extension = determine_file_type(file_path.name)
|
95 |
+
if file_extension in extension_to_parser:
|
96 |
+
parsed_contents[file_path.name] = extension_to_parser[file_extension](file_path.name)
|
97 |
+
else:
|
98 |
+
parsed_contents[file_path.name] = f"Unsupported file type: {file_extension}"
|
99 |
+
|
100 |
+
filename_end = get_file_path_end(file_path.name)
|
101 |
+
|
102 |
+
file_names.append(filename_end)
|
103 |
+
|
104 |
+
return parsed_contents, file_names
|
105 |
+
|
106 |
+
def text_regex_clean(text):
|
107 |
+
# Merge hyphenated words
|
108 |
+
text = re.sub(r"(\w+)-\n(\w+)", r"\1\2", text)
|
109 |
+
# If a double newline ends in a letter, add a full stop.
|
110 |
+
text = re.sub(r'(?<=[a-zA-Z])\n\n', '.\n\n', text)
|
111 |
+
# Fix newlines in the middle of sentences
|
112 |
+
text = re.sub(r"(?<!\n\s)\n(?!\s\n)", " ", text.strip())
|
113 |
+
# Remove multiple newlines
|
114 |
+
text = re.sub(r"\n\s*\n", "\n\n", text)
|
115 |
+
text = re.sub(r" ", " ", text)
|
116 |
+
# Add full stops and new lines between words with no space between where the second one has a capital letter
|
117 |
+
text = re.sub(r'(?<=[a-z])(?=[A-Z])', '. \n\n', text)
|
118 |
+
|
119 |
+
return text
|
120 |
+
|
121 |
+
def parse_csv_or_excel(file_paths, text_column = "text"):
|
122 |
+
"""
|
123 |
+
Read in a CSV or Excel file.
|
124 |
+
|
125 |
+
Parameters:
|
126 |
+
file_path (str): Path to the CSV file.
|
127 |
+
text_column (str): Name of the column in the CSV file that contains the text content.
|
128 |
+
|
129 |
+
Returns:
|
130 |
+
Pandas DataFrame: Dataframe output from file read
|
131 |
+
"""
|
132 |
+
|
133 |
+
file_names = []
|
134 |
+
out_df = pd.DataFrame()
|
135 |
+
|
136 |
+
for file_path in file_paths:
|
137 |
+
file_extension = determine_file_type(file_path.name)
|
138 |
+
file_name = get_file_path_end(file_path.name)
|
139 |
+
|
140 |
+
if file_extension == ".csv":
|
141 |
+
df = pd.read_csv(file_path.name)
|
142 |
+
if text_column not in df.columns: return pd.DataFrame(), ['Please choose a valid column name']
|
143 |
+
df['source'] = file_name
|
144 |
+
df['page_section'] = ""
|
145 |
+
elif file_extension == ".xlsx":
|
146 |
+
df = pd.read_excel(file_path.name, engine='openpyxl')
|
147 |
+
if text_column not in df.columns: return pd.DataFrame(), ['Please choose a valid column name']
|
148 |
+
df['source'] = file_name
|
149 |
+
df['page_section'] = ""
|
150 |
+
else:
|
151 |
+
print(f"Unsupported file type: {file_extension}")
|
152 |
+
return pd.DataFrame(), ['Please choose a valid file type']
|
153 |
+
|
154 |
+
file_names.append(file_name)
|
155 |
+
out_df = pd.concat([out_df, df])
|
156 |
+
|
157 |
+
#if text_column not in df.columns:
|
158 |
+
# return f"Column '{text_column}' not found in {file_path}"
|
159 |
+
#text_out = " ".join(df[text_column].dropna().astype(str))
|
160 |
+
return out_df, file_names
|
161 |
+
|
162 |
+
def parse_excel(file_path, text_column):
|
163 |
+
"""
|
164 |
+
Read text from an Excel file.
|
165 |
+
|
166 |
+
Parameters:
|
167 |
+
file_path (str): Path to the Excel file.
|
168 |
+
text_column (str): Name of the column in the Excel file that contains the text content.
|
169 |
+
|
170 |
+
Returns:
|
171 |
+
Pandas DataFrame: Dataframe output from file read
|
172 |
+
"""
|
173 |
+
df = pd.read_excel(file_path, engine='openpyxl')
|
174 |
+
#if text_column not in df.columns:
|
175 |
+
# return f"Column '{text_column}' not found in {file_path}"
|
176 |
+
#text_out = " ".join(df[text_column].dropna().astype(str))
|
177 |
+
return df
|
178 |
+
|
179 |
+
def parse_pdf(file) -> List[str]:
|
180 |
+
|
181 |
+
"""
|
182 |
+
Extract text from a PDF file.
|
183 |
+
|
184 |
+
Parameters:
|
185 |
+
file_path (str): Path to the PDF file.
|
186 |
+
|
187 |
+
Returns:
|
188 |
+
List[str]: Extracted text from the PDF.
|
189 |
+
"""
|
190 |
+
|
191 |
+
output = []
|
192 |
+
#for file in files:
|
193 |
+
print(file) # .name
|
194 |
+
pdf = PdfReader(file) #[i] .name[i]
|
195 |
+
|
196 |
+
for page in pdf.pages:
|
197 |
+
text = page.extract_text()
|
198 |
+
|
199 |
+
text = text_regex_clean(text)
|
200 |
+
|
201 |
+
output.append(text)
|
202 |
+
return output
|
203 |
+
|
204 |
+
def parse_docx(file_path):
|
205 |
+
"""
|
206 |
+
Reads the content of a .docx file and returns it as a string.
|
207 |
+
|
208 |
+
Parameters:
|
209 |
+
- file_path (str): Path to the .docx file.
|
210 |
+
|
211 |
+
Returns:
|
212 |
+
- str: Content of the .docx file.
|
213 |
+
"""
|
214 |
+
doc = Doc(file_path)
|
215 |
+
full_text = []
|
216 |
+
for para in doc.paragraphs:
|
217 |
+
para = text_regex_clean(para)
|
218 |
+
|
219 |
+
full_text.append(para.text.replace(" ", " ").strip())
|
220 |
+
return '\n'.join(full_text)
|
221 |
+
|
222 |
+
def parse_txt(file_path):
|
223 |
+
"""
|
224 |
+
Read text from a TXT or HTML file.
|
225 |
+
|
226 |
+
Parameters:
|
227 |
+
file_path (str): Path to the TXT or HTML file.
|
228 |
+
|
229 |
+
Returns:
|
230 |
+
str: Text content of the file.
|
231 |
+
"""
|
232 |
+
with open(file_path, 'r', encoding="utf-8") as file:
|
233 |
+
file_contents = file.read().replace(" ", " ").strip()
|
234 |
+
|
235 |
+
file_contents = text_regex_clean(file_contents)
|
236 |
+
|
237 |
+
return file_contents
|
238 |
+
|
239 |
+
def parse_html(page_url, div_filter="p"):
|
240 |
+
"""
|
241 |
+
Determine if the source is a web URL or a local HTML file, extract the content based on the div of choice. Also tries to extract dates (WIP)
|
242 |
+
|
243 |
+
Parameters:
|
244 |
+
page_url (str): The web URL or local file path.
|
245 |
+
|
246 |
+
Returns:
|
247 |
+
str: Extracted content.
|
248 |
+
"""
|
249 |
+
|
250 |
+
def is_web_url(s):
|
251 |
+
"""
|
252 |
+
Check if the input string is a web URL.
|
253 |
+
"""
|
254 |
+
return s.startswith("http://") or s.startswith("https://")
|
255 |
+
|
256 |
+
def is_local_html_file(s):
|
257 |
+
"""
|
258 |
+
Check if the input string is a path to a local HTML file.
|
259 |
+
"""
|
260 |
+
return (s.endswith(".html") or s.endswith(".htm")) and os.path.isfile(s)
|
261 |
+
|
262 |
+
def extract_text_from_source(source):
|
263 |
+
"""
|
264 |
+
Determine if the source is a web URL or a local HTML file,
|
265 |
+
and then extract its content accordingly.
|
266 |
+
|
267 |
+
Parameters:
|
268 |
+
source (str): The web URL or local file path.
|
269 |
+
|
270 |
+
Returns:
|
271 |
+
str: Extracted content.
|
272 |
+
"""
|
273 |
+
if is_web_url(source):
|
274 |
+
response = requests.get(source)
|
275 |
+
response.raise_for_status() # Raise an HTTPError for bad responses
|
276 |
+
return response.text.replace(" ", " ").strip()
|
277 |
+
elif is_local_html_file(source):
|
278 |
+
with open(source, 'r', encoding='utf-8') as file:
|
279 |
+
file_out = file.read().replace
|
280 |
+
return file_out
|
281 |
+
else:
|
282 |
+
raise ValueError("Input is neither a valid web URL nor a local HTML file path.")
|
283 |
+
|
284 |
+
|
285 |
+
def clean_html_data(data, date_filter="", div_filt="p"):
|
286 |
+
"""
|
287 |
+
Extracts and cleans data from HTML content.
|
288 |
+
|
289 |
+
Parameters:
|
290 |
+
data (str): HTML content to be parsed.
|
291 |
+
date_filter (str, optional): Date string to filter results. If set, only content with a date greater than this will be returned.
|
292 |
+
div_filt (str, optional): HTML tag to search for text content. Defaults to "p".
|
293 |
+
|
294 |
+
Returns:
|
295 |
+
tuple: Contains extracted text and date as strings. Returns empty strings if not found.
|
296 |
+
"""
|
297 |
+
|
298 |
+
soup = BeautifulSoup(data, 'html.parser')
|
299 |
+
|
300 |
+
# Function to exclude div with id "bar"
|
301 |
+
def exclude_div_with_id_bar(tag):
|
302 |
+
return tag.has_attr('id') and tag['id'] == 'related-links'
|
303 |
+
|
304 |
+
text_elements = soup.find_all(div_filt)
|
305 |
+
date_elements = soup.find_all(div_filt, {"class": "page-neutral-intro__meta"})
|
306 |
+
|
307 |
+
# Extract date
|
308 |
+
date_out = ""
|
309 |
+
if date_elements:
|
310 |
+
date_out = re.search(">(.*?)<", str(date_elements[0])).group(1)
|
311 |
+
date_dt = dateutil.parser.parse(date_out)
|
312 |
+
|
313 |
+
if date_filter:
|
314 |
+
date_filter_dt = dateutil.parser.parse(date_filter)
|
315 |
+
if date_dt < date_filter_dt:
|
316 |
+
return '', date_out
|
317 |
+
|
318 |
+
# Extract text
|
319 |
+
text_out_final = ""
|
320 |
+
if text_elements:
|
321 |
+
text_out_final = '\n'.join(paragraph.text for paragraph in text_elements)
|
322 |
+
text_out_final = text_regex_clean(text_out_final)
|
323 |
+
else:
|
324 |
+
print(f"No elements found with tag '{div_filt}'. No text returned.")
|
325 |
+
|
326 |
+
return text_out_final, date_out
|
327 |
+
|
328 |
+
|
329 |
+
#page_url = "https://pypi.org/project/InstructorEmbedding/" #'https://www.ons.gov.uk/visualisations/censusareachanges/E09000022/index.html'
|
330 |
+
|
331 |
+
html_text = extract_text_from_source(page_url)
|
332 |
+
#print(page.text)
|
333 |
+
|
334 |
+
texts = []
|
335 |
+
metadatas = []
|
336 |
+
|
337 |
+
clean_text, date = clean_html_data(html_text, date_filter="", div_filt=div_filter)
|
338 |
+
texts.append(clean_text)
|
339 |
+
metadatas.append({"source": page_url, "date":str(date)})
|
340 |
+
|
341 |
+
#print(metadatas)
|
342 |
+
|
343 |
+
return texts, metadatas, page_url
|
344 |
+
|
345 |
+
def get_file_path_end(file_path):
|
346 |
+
match = re.search(r'(.*[\/\\])?(.+)$', file_path)
|
347 |
+
|
348 |
+
filename_end = match.group(2) if match else ''
|
349 |
+
|
350 |
+
return filename_end
|
351 |
+
|
352 |
+
# +
|
353 |
+
# Convert parsed text to docs
|
354 |
+
# -
|
355 |
+
|
356 |
+
def text_to_docs(text_dict: dict, chunk_size: int = chunk_size) -> List[Document]:
|
357 |
+
"""
|
358 |
+
Converts the output of parse_file (a dictionary of file paths to content)
|
359 |
+
to a list of Documents with metadata.
|
360 |
+
"""
|
361 |
+
|
362 |
+
doc_sections = []
|
363 |
+
parent_doc_sections = []
|
364 |
+
|
365 |
+
for file_path, content in text_dict.items():
|
366 |
+
ext = os.path.splitext(file_path)[1].lower()
|
367 |
+
|
368 |
+
# Depending on the file extension, handle the content
|
369 |
+
if ext == '.pdf':
|
370 |
+
docs, page_docs = pdf_text_to_docs(content, chunk_size)
|
371 |
+
elif ext in ['.html', '.htm', '.txt', '.docx']:
|
372 |
+
docs = html_text_to_docs(content, chunk_size)
|
373 |
+
elif ext in ['.csv', '.xlsx']:
|
374 |
+
docs, page_docs = csv_excel_text_to_docs(content, chunk_size)
|
375 |
+
else:
|
376 |
+
print(f"Unsupported file type {ext} for {file_path}. Skipping.")
|
377 |
+
continue
|
378 |
+
|
379 |
+
|
380 |
+
filename_end = get_file_path_end(file_path)
|
381 |
+
|
382 |
+
#match = re.search(r'(.*[\/\\])?(.+)$', file_path)
|
383 |
+
#filename_end = match.group(2) if match else ''
|
384 |
+
|
385 |
+
# Add filename as metadata
|
386 |
+
for doc in docs: doc.metadata["source"] = filename_end
|
387 |
+
#for parent_doc in parent_docs: parent_doc.metadata["source"] = filename_end
|
388 |
+
|
389 |
+
doc_sections.extend(docs)
|
390 |
+
#parent_doc_sections.extend(parent_docs)
|
391 |
+
|
392 |
+
return doc_sections#, page_docs
|
393 |
+
|
394 |
+
def pdf_text_to_docs(text, chunk_size: int = chunk_size) -> List[Document]:
|
395 |
+
"""Converts a string or list of strings to a list of Documents
|
396 |
+
with metadata."""
|
397 |
+
|
398 |
+
#print(text)
|
399 |
+
|
400 |
+
if isinstance(text, str):
|
401 |
+
# Take a single string as one page
|
402 |
+
text = [text]
|
403 |
+
|
404 |
+
page_docs = [Document(page_content=page, metadata={"page": page}) for page in text]
|
405 |
+
|
406 |
+
|
407 |
+
# Add page numbers as metadata
|
408 |
+
for i, doc in enumerate(page_docs):
|
409 |
+
doc.metadata["page"] = i + 1
|
410 |
+
|
411 |
+
print("page docs are: ")
|
412 |
+
print(page_docs)
|
413 |
+
|
414 |
+
# Split pages into sections
|
415 |
+
doc_sections = []
|
416 |
+
|
417 |
+
for doc in page_docs:
|
418 |
+
|
419 |
+
#print("page content: ")
|
420 |
+
#print(doc.page_content)
|
421 |
+
|
422 |
+
if doc.page_content == '':
|
423 |
+
sections = ['']
|
424 |
+
|
425 |
+
else:
|
426 |
+
text_splitter = RecursiveCharacterTextSplitter(
|
427 |
+
chunk_size=chunk_size,
|
428 |
+
separators=split_strat,#["\n\n", "\n", ".", "!", "?", ",", " ", ""],
|
429 |
+
chunk_overlap=chunk_overlap,
|
430 |
+
add_start_index=True
|
431 |
+
)
|
432 |
+
sections = text_splitter.split_text(doc.page_content)
|
433 |
+
|
434 |
+
for i, section in enumerate(sections):
|
435 |
+
doc = Document(
|
436 |
+
page_content=section, metadata={"page": doc.metadata["page"], "section": i, "page_section": f"{doc.metadata['page']}-{i}"})
|
437 |
+
|
438 |
+
|
439 |
+
doc_sections.append(doc)
|
440 |
+
|
441 |
+
return doc_sections, page_docs#, parent_doc
|
442 |
+
|
443 |
+
def html_text_to_docs(texts, metadatas, chunk_size:int = chunk_size):
|
444 |
+
|
445 |
+
text_splitter = RecursiveCharacterTextSplitter(
|
446 |
+
separators=split_strat,#["\n\n", "\n", ".", "!", "?", ",", " ", ""],
|
447 |
+
chunk_size=chunk_size,
|
448 |
+
chunk_overlap=chunk_overlap,
|
449 |
+
length_function=len,
|
450 |
+
add_start_index=True
|
451 |
+
)
|
452 |
+
|
453 |
+
#print(texts)
|
454 |
+
#print(metadatas)
|
455 |
+
|
456 |
+
documents = text_splitter.create_documents(texts, metadatas=metadatas)
|
457 |
+
|
458 |
+
for i, section in enumerate(documents):
|
459 |
+
section.metadata["page_section"] = i + 1
|
460 |
+
|
461 |
+
|
462 |
+
|
463 |
+
return documents
|
464 |
+
|
465 |
+
def write_out_metadata_as_string(metadata_in):
|
466 |
+
# If metadata_in is a single dictionary, wrap it in a list
|
467 |
+
if isinstance(metadata_in, dict):
|
468 |
+
metadata_in = [metadata_in]
|
469 |
+
|
470 |
+
metadata_string = [f"{' '.join(f'{k}: {v}' for k, v in d.items() if k != 'page_section')}" for d in metadata_in] # ['metadata']
|
471 |
+
return metadata_string
|
472 |
+
|
473 |
+
def csv_excel_text_to_docs(df, text_column='text', chunk_size=None) -> List[Document]:
|
474 |
+
"""Converts a DataFrame's content to a list of Documents with metadata."""
|
475 |
+
|
476 |
+
doc_sections = []
|
477 |
+
df[text_column] = df[text_column].astype(str) # Ensure column is a string column
|
478 |
+
|
479 |
+
# For each row in the dataframe
|
480 |
+
for idx, row in df.iterrows():
|
481 |
+
# Extract the text content for the document
|
482 |
+
doc_content = row[text_column]
|
483 |
+
|
484 |
+
# Generate metadata containing other columns' data
|
485 |
+
metadata = {"row": idx + 1}
|
486 |
+
for col, value in row.items():
|
487 |
+
if col != text_column:
|
488 |
+
metadata[col] = value
|
489 |
+
|
490 |
+
metadata_string = write_out_metadata_as_string(metadata)[0]
|
491 |
+
|
492 |
+
|
493 |
+
|
494 |
+
# If chunk_size is provided, split the text into chunks
|
495 |
+
if chunk_size:
|
496 |
+
# Assuming you have a text splitter function similar to the PDF handling
|
497 |
+
text_splitter = RecursiveCharacterTextSplitter(
|
498 |
+
chunk_size=chunk_size,
|
499 |
+
# Other arguments as required by the splitter
|
500 |
+
)
|
501 |
+
sections = text_splitter.split_text(doc_content)
|
502 |
+
|
503 |
+
|
504 |
+
# For each section, create a Document object
|
505 |
+
for i, section in enumerate(sections):
|
506 |
+
section = '. '.join([metadata_string, section])
|
507 |
+
doc = Document(page_content=section,
|
508 |
+
metadata={**metadata, "section": i, "row_section": f"{metadata['row']}-{i}"})
|
509 |
+
doc_sections.append(doc)
|
510 |
+
else:
|
511 |
+
# If no chunk_size is provided, create a single Document object for the row
|
512 |
+
doc_content = '. '.join([metadata_string, doc_content])
|
513 |
+
doc = Document(page_content=doc_content, metadata=metadata)
|
514 |
+
doc_sections.append(doc)
|
515 |
+
|
516 |
+
return doc_sections
|
517 |
+
|
518 |
+
# # Functions for working with documents after loading them back in
|
519 |
+
|
520 |
+
def pull_out_data(series):
|
521 |
+
|
522 |
+
# define a lambda function to convert each string into a tuple
|
523 |
+
to_tuple = lambda x: eval(x)
|
524 |
+
|
525 |
+
# apply the lambda function to each element of the series
|
526 |
+
series_tup = series.apply(to_tuple)
|
527 |
+
|
528 |
+
series_tup_content = list(zip(*series_tup))[1]
|
529 |
+
|
530 |
+
series = pd.Series(list(series_tup_content))#.str.replace("^Main post content", "", regex=True).str.strip()
|
531 |
+
|
532 |
+
return series
|
533 |
+
|
534 |
+
def docs_from_csv(df):
|
535 |
+
|
536 |
+
import ast
|
537 |
+
|
538 |
+
documents = []
|
539 |
+
|
540 |
+
page_content = pull_out_data(df["0"])
|
541 |
+
metadatas = pull_out_data(df["1"])
|
542 |
+
|
543 |
+
for x in range(0,len(df)):
|
544 |
+
new_doc = Document(page_content=page_content[x], metadata=metadatas[x])
|
545 |
+
documents.append(new_doc)
|
546 |
+
|
547 |
+
return documents
|
548 |
+
|
549 |
+
def docs_from_lists(docs, metadatas):
|
550 |
+
|
551 |
+
documents = []
|
552 |
+
|
553 |
+
for x, doc in enumerate(docs):
|
554 |
+
new_doc = Document(page_content=doc, metadata=metadatas[x])
|
555 |
+
documents.append(new_doc)
|
556 |
+
|
557 |
+
return documents
|
558 |
+
|
559 |
+
def docs_elements_from_csv_save(docs_path="documents.csv"):
|
560 |
+
|
561 |
+
documents = pd.read_csv(docs_path)
|
562 |
+
|
563 |
+
docs_out = docs_from_csv(documents)
|
564 |
+
|
565 |
+
out_df = pd.DataFrame(docs_out)
|
566 |
+
|
567 |
+
docs_content = pull_out_data(out_df[0].astype(str))
|
568 |
+
|
569 |
+
docs_meta = pull_out_data(out_df[1].astype(str))
|
570 |
+
|
571 |
+
doc_sources = [d['source'] for d in docs_meta]
|
572 |
+
|
573 |
+
return out_df, docs_content, docs_meta, doc_sources
|
574 |
+
|
575 |
+
# ## Create embeddings and save faiss vector store to the path specified in `save_to`
|
576 |
+
|
577 |
+
def load_embeddings(model_name = "BAAI/bge-base-en-v1.5"):
|
578 |
+
|
579 |
+
#if model_name == "hkunlp/instructor-large":
|
580 |
+
# embeddings_func = HuggingFaceInstructEmbeddings(model_name=model_name,
|
581 |
+
# embed_instruction="Represent the paragraph for retrieval: ",
|
582 |
+
# query_instruction="Represent the question for retrieving supporting documents: "
|
583 |
+
# )
|
584 |
+
|
585 |
+
#else:
|
586 |
+
embeddings_func = HuggingFaceEmbeddings(model_name=model_name)
|
587 |
+
|
588 |
+
global embeddings
|
589 |
+
|
590 |
+
embeddings = embeddings_func
|
591 |
+
|
592 |
+
return embeddings_func
|
593 |
+
|
594 |
+
def embed_faiss_save_to_zip(docs_out, save_to="faiss_lambeth_census_embedding", model_name = "BAAI/bge-base-en-v1.5"):
|
595 |
+
|
596 |
+
load_embeddings(model_name=model_name)
|
597 |
+
|
598 |
+
#embeddings_fast = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")
|
599 |
+
|
600 |
+
print(f"> Total split documents: {len(docs_out)}")
|
601 |
+
|
602 |
+
vectorstore = FAISS.from_documents(documents=docs_out, embedding=embeddings)
|
603 |
+
|
604 |
+
|
605 |
+
if Path(save_to).exists():
|
606 |
+
vectorstore.save_local(folder_path=save_to)
|
607 |
+
|
608 |
+
print("> DONE")
|
609 |
+
print(f"> Saved to: {save_to}")
|
610 |
+
|
611 |
+
### Save as zip, then remove faiss/pkl files to allow for upload to huggingface
|
612 |
+
|
613 |
+
import shutil
|
614 |
+
|
615 |
+
shutil.make_archive(save_to, 'zip', save_to)
|
616 |
+
|
617 |
+
os.remove(save_to + "/index.faiss")
|
618 |
+
os.remove(save_to + "/index.pkl")
|
619 |
+
|
620 |
+
shutil.move(save_to + '.zip', save_to + "/" + save_to + '.zip')
|
621 |
+
|
622 |
+
return vectorstore
|
623 |
+
|
624 |
+
def docs_to_chroma_save(embeddings, docs_out:PandasDataFrame, save_to:str):
|
625 |
+
print(f"> Total split documents: {len(docs_out)}")
|
626 |
+
|
627 |
+
vectordb = Chroma.from_documents(documents=docs_out,
|
628 |
+
embedding=embeddings,
|
629 |
+
persist_directory=save_to)
|
630 |
+
|
631 |
+
# persiste the db to disk
|
632 |
+
vectordb.persist()
|
633 |
+
|
634 |
+
print("> DONE")
|
635 |
+
print(f"> Saved to: {save_to}")
|
636 |
+
|
637 |
+
return vectordb
|
638 |
+
|
639 |
+
def sim_search_local_saved_vec(query, k_val, save_to="faiss_lambeth_census_embedding"):
|
640 |
+
|
641 |
+
load_embeddings()
|
642 |
+
|
643 |
+
docsearch = FAISS.load_local(folder_path=save_to, embeddings=embeddings)
|
644 |
+
|
645 |
+
|
646 |
+
display(Markdown(question))
|
647 |
+
|
648 |
+
search = docsearch.similarity_search_with_score(query, k=k_val)
|
649 |
+
|
650 |
+
for item in search:
|
651 |
+
print(item[0].page_content)
|
652 |
+
print(f"Page: {item[0].metadata['source']}")
|
653 |
+
print(f"Date: {item[0].metadata['date']}")
|
654 |
+
print(f"Score: {item[1]}")
|
655 |
+
print("---")
|
chatfuncs/ingest_borough_plan.py
ADDED
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import ingest as ing
|
2 |
+
|
3 |
+
borough_plan_text, file_names = ing.parse_file([open("Lambeth_2030-Our_Future_Our_Lambeth.pdf")])
|
4 |
+
print("Borough plan text created")
|
5 |
+
|
6 |
+
print(borough_plan_text)
|
7 |
+
|
8 |
+
borough_plan_docs = ing.text_to_docs(borough_plan_text)
|
9 |
+
print("Borough plan docs created")
|
10 |
+
|
11 |
+
embedding_model = "BAAI/bge-base-en-v1.5"
|
12 |
+
|
13 |
+
embeddings = ing.load_embeddings(model_name = embedding_model)
|
14 |
+
ing.embed_faiss_save_to_zip(borough_plan_docs, save_to="faiss_embedding", model_name = embedding_model)
|