utils.py

# LLM
# Ollama for local tests
from langchain.llms import Ollama
from langchain.callbacks.manager import CallbackManager
from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
from langchain.llms import Ollama
# Ref.: https://mistral.ai/news/mixtral-of-experts/#instructed-models
# Q5_K_M quantzation flavor for best quality/recommended tradeoff (memory is no problem here)
# Ref.: https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-GGUF#provided-files
MISTRAL = "mistral:7b-instruct-v0.2-q5_K_M"
# Q4_K quantization flavor for best memory/quality/recommended tradeoff
# Ref.: https://huggingface.co/TheBloke/Mixtral-8x7B-Instruct-v0.1-GGUF#provided-files
# mixtral:8x7b-instruct-v0.1-q4_K_M was sadly still too big for my Mac
MIXTRAL = "mixtral:8x7b-instruct-v0.1-q3_K_L"
# Llama2 13B 
# Ref.: https://huggingface.co/TheBloke/Llama-2-13B-GGUF
LLAMA2 = "llama2:13b-chat-q5_K_M"
mistral = Ollama(
    model=MISTRAL,
    callback_manager=CallbackManager([StreamingStdOutCallbackHandler()]),
    # Ref.: https://api.python.langchain.com/en/latest/llms/langchain_community.llms.ollama.Ollama.html#langchain_community.llms.ollama.Ollama.format
    # format="json"
)
mixtral = Ollama(
    model=MIXTRAL,
    callback_manager=CallbackManager([StreamingStdOutCallbackHandler()])
)
llama2 = Ollama(
    model=LLAMA2,
    callback_manager=CallbackManager([StreamingStdOutCallbackHandler()])
)


# LOAD
from langchain_community.document_loaders import UnstructuredMarkdownLoader
from langchain_community.document_loaders import TextLoader
from langchain_community.document_loaders import PyPDFLoader

FILES = {
    'md': [
        # "Présentation modes dégradés-20230120_112423-Enregistrement de la réunion.md",
        "YouTube - Mode secours telephonie.md"
    ],
    'pdf': [
        # "SI-Samu_Fiche procédure_Mode dégradé_Perte de CRRA.pdf",
        # "[SI-Samu] Fiche mémo - Procédure Mode dégradé.pdf",
        "SI-Samu_Documentation_produit_SF4_J18HF2_20231219 - mode secours seul.pdf",
        # "SI-Samu_Documentation_produit_SF4_J18HF2_20231219.pdf"
    ]
}

def load_data(files):
    data = {'md': [], 'pdf': []}
    for pdf in files['pdf']:
        data['pdf'].extend(PyPDFLoader('resources/' + pdf).load())
    for md in files['md']:
        data['md'].extend(TextLoader('resources/' + md).load())
    return data

def to_full_data(data):
    return [
        *data['md'],
        *data['pdf']
    ]

# SPLIT
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.text_splitter import MarkdownHeaderTextSplitter

def split_MD_then_recursive(data):
    # - First use MarkDown title splitter on .MD and then RecursiveSplitter on all
    # MD splits
    markdown_splitter = MarkdownHeaderTextSplitter(headers_to_split_on=[
        ("#", "Titre 1"),
        ("##", "Titre 2"),
        ("###", "Titre 3"),
    ], strip_headers=False)
    md_header_splits = data['pdf'].copy()
    for md in data['md']:
        md_header_splits.extend(markdown_splitter.split_text(md.page_content))

    # Char-level splits
    text_splitter=RecursiveCharacterTextSplitter(
        chunk_size=500, 
        chunk_overlap=50  # to improve results quality
    )
    # Split
    return text_splitter.split_documents(md_header_splits)

# EMBED
# Directly done in the different scripts

# RETRIEVE
from langchain.storage import InMemoryStore
from langchain.retrievers import ParentDocumentRetriever, BM25Retriever, EnsembleRetriever
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.vectorstores import Chroma

# Ensemble is based on weight fusion (Reciprocal Rank Fusion) | Ref.: https://safjan.com/implementing-rank-fusion-in-python/
def get_parent_ensemble_retriever(embeddings, full_data, all_splits, k=4, parent_chunk_size=2000, child_chunk_size=400, collection_name="store"):
    # - ParentDocumentRetriever: embed small chunks but retrieve with bigger context
    # This text splitter is used to create the parent documents
    parent_splitter = RecursiveCharacterTextSplitter(chunk_size=parent_chunk_size)
    # This text splitter is used to create the child documents
    # It should create documents smaller than the parent (don't make bigger than 512 as most embeddings trunk after that)
    child_splitter = RecursiveCharacterTextSplitter(chunk_size=child_chunk_size)
    # The vectorstore to use to index the child chunks
    parent_vectorstore = Chroma(
        collection_name=collection_name, 
        embedding_function=embeddings
    )
    # The storage layer for the parent documents
    parent_store = InMemoryStore()
    parent_retriever = ParentDocumentRetriever(
        vectorstore=parent_vectorstore,
        docstore=parent_store,
        child_splitter=child_splitter,
        parent_splitter=parent_splitter,
        search_kwargs={
            "k": k,
            # "score_threshold": 0.5
        },
        # search_type="mmr"
    )
    parent_retriever.add_documents(full_data)

    # - EnsembleRetriever
    # BM25 logic
    bm25_retriever = BM25Retriever.from_texts(
        list(map(lambda s: s.page_content, all_splits)), 
        metadatas=list(map(lambda s: {"retriever": "BM25 sparse similiarity", **s.metadata}, all_splits))
        
    )
    bm25_retriever.k = k

    # Ensemble of BM25 + vectorstore on parent retriever
    return EnsembleRetriever(
        retrievers=[parent_retriever, bm25_retriever], weights=[0.5, 0.5]
    )

# PROMPT
# Add more context to query + update system prompt to make it speak French
# Ref.: https://stackoverflow.com/questions/76554411/unable-to-pass-prompt-template-to-retrievalqa-in-langchain
# Ref.: https://community.openai.com/t/how-to-prevent-chatgpt-from-answering-questions-that-are-outside-the-scope-of-the-provided-context-in-the-system-role-message/112027/7
from langchain import PromptTemplate
template = """
System: You are helping a user of "bandeau téléphonique SI-SAMU" (a CTI - Computer Telephony Integration - system) during system failure as he needs to use its local backup phone.
Context information is below. Given the context information and not prior knowledge, answer the query.
Language: Answer in French and using "vous".
---
Context: {context}
---
Question: {question}
---
Réponse :
"""
PROMPT = PromptTemplate(template=template, input_variables=['question', 'context'])

# RESULTS
class color:
   PURPLE = '\033[95m'
   CYAN = '\033[96m'
   DARKCYAN = '\033[36m'
   BLUE = '\033[94m'
   GREEN = '\033[92m'
   YELLOW = '\033[93m'
   RED = '\033[91m'
   BOLD = '\033[1m'
   UNDERLINE = '\033[4m'
   END = '\033[0m'

def parse_answer(answer):
    print(f">> {answer['query']}")
    print(f">> {answer['result']}")
    print(">> Sources :")
    for doc in answer['source_documents']:
        page = ''
        if 'page' in doc.metadata:
            page = f" (page {doc.metadata['page']})"
        source = ''
        if 'source' in doc.metadata:
            source = doc.metadata['source']
        titles = ['Titre 1', 'Titre 2', 'Titre 3']
        for title in titles:
            if title in doc.metadata:
                source += f" > {doc.metadata[title]}"
        retriever = f"B25" if 'retriever' in doc.metadata else "vectorstore"
        print(f">>> {color.BOLD}{source}{page} [{retriever}]{color.END}: {doc.page_content}\n---")
    print("--------\n\n")