demo

app.py

@@ -38,10 +38,10 @@ import requests
 import os
 import gradio as gr
 from sentence_transformers import SentenceTransformer, CrossEncoder, util
-#from torch import tensor as torch_tensor
-#from datasets import load_dataset
+from torch import tensor as torch_tensor
+from datasets import load_dataset
 
-from greg_funcs import mrkl_rspnd, get_cot
+from greg_funcs import mrkl_rspnd
 
 
 OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")  
@@ -397,7 +397,6 @@ class ChatWrapper:
 
 
             response = mrkl_rspnd(inp)
-            cot = get_cot(response)
             output = response['output']
 
             """
@@ -464,7 +463,7 @@ class ChatWrapper:
             raise e
         finally:
             self.lock.release()
-        return history, history, html_video, temp_file, html_audio, temp_aud_file, cot
+        return history, history, html_video, temp_file, html_audio, temp_aud_file, ""
         # return history, history, html_audio, temp_aud_file, ""
 
 
@@ -656,8 +655,7 @@ with gr.Blocks(css=CSS) as block:
 
             with gr.Column(scale=7):
                 chatbot = gr.Chatbot()
-        with gr.Accordion(label='Show AI chain of thought: ', open=False,):
-            ai_cot = gr.HTML(show_label=False)
+
         with gr.Row():
             message = gr.Textbox(label="What's on your mind??",
                                  placeholder=PLACEHOLDER,
@@ -854,7 +852,7 @@ with gr.Blocks(css=CSS) as block:
                                  anticipation_level_state, joy_level_state, trust_level_state, fear_level_state,
                                  surprise_level_state, sadness_level_state, disgust_level_state, anger_level_state,
                                  lang_level_state, translate_to_state, literary_style_state],
-                   outputs=[chatbot, history_state, video_html, my_file, audio_html, tmp_aud_file, ai_cot])
+                   outputs=[chatbot, history_state, video_html, my_file, audio_html, tmp_aud_file, message])
     # outputs=[chatbot, history_state, audio_html, tmp_aud_file, message])
 
     submit.click(chat, inputs=[message, history_state, chain_state, trace_chain_state,
@@ -863,8 +861,8 @@ with gr.Blocks(css=CSS) as block:
                                anticipation_level_state, joy_level_state, trust_level_state, fear_level_state,
                                surprise_level_state, sadness_level_state, disgust_level_state, anger_level_state,
                                lang_level_state, translate_to_state, literary_style_state],
-                 outputs=[chatbot, history_state, video_html, my_file, audio_html, tmp_aud_file, ai_cot])
-        # history, history, html_video, temp_file, html_audio, temp_aud_file, cot
+                 outputs=[chatbot, history_state, video_html, my_file, audio_html, tmp_aud_file, message])
+    # outputs=[chatbot, history_state, audio_html, tmp_aud_file, message])
 
 
 block.launch(debug=True)

data/mychromadb/chroma-collections.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:f53c5c7bfaf512515b19c15daf41df9f862138f1fad64bfa27490123a6ae0630
-size 592

data/mychromadb/chroma-embeddings.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:00c915a2802e985b9b5a95b9c90d1c541cea4b671e64ac2b557e9f4dec0c9648
-size 3352718

data/mychromadb/index/id_to_uuid_aca62790-d606-4764-91d9-08324ea54984.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:8a21d750f00302ae6e2e80f188dd5910f14cdc773dcfe8cc420a9b226b1a06ed
-size 44935

data/mychromadb/index/id_to_uuid_f75ac8ed-ecb5-4656-84d1-90e84f6f083a.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:39ba580b21452d0c0363b9621737aa4f819cb6e8d811888905aeb08c1370dd46
-size 24704

data/mychromadb/index/index_aca62790-d606-4764-91d9-08324ea54984.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:1dd287418a49ded3f5e07a60872c421073d37942ff0b2c43cd69cdccdd069040
-size 2341688

data/mychromadb/index/index_f75ac8ed-ecb5-4656-84d1-90e84f6f083a.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:eb8680a43caa02daca0e4fd5ce8abadab705b2ba994bdb3d056164406c606d00
-size 1292260

data/mychromadb/index/index_metadata_aca62790-d606-4764-91d9-08324ea54984.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:92cfe00d3fcdf3bffa64e13342d5210490b113a67a82fe54460622777e35bf2b
-size 74

data/mychromadb/index/index_metadata_f75ac8ed-ecb5-4656-84d1-90e84f6f083a.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:505fa51457f99d2d4bdf6860f2326abc7d2d45f6f0462657b6de1aa564945472
-size 74

data/mychromadb/index/uuid_to_id_aca62790-d606-4764-91d9-08324ea54984.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:dfa0915ece0f552a799dbce3e57e44fe612eb118ab09fe17ab71b98ab1a71a32
-size 52582

data/mychromadb/index/uuid_to_id_f75ac8ed-ecb5-4656-84d1-90e84f6f083a.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:0cfd370f60ead46494e5acd164bccd9dc2a1808e498a98bded54d94dedd73085
-size 28906

greg_funcs.py

@@ -1,6 +1,6 @@
 from sentence_transformers import SentenceTransformer, CrossEncoder, util
-#from torch import tensor as torch_tensor
-#from datasets import load_dataset
+from torch import tensor as torch_tensor
+from datasets import load_dataset
 
 from langchain.llms import OpenAI
 from langchain.docstore.document import Document
@@ -13,7 +13,6 @@ from langchain.agents import initialize_agent, Tool
 import sqlite3
 #import pandas as pd
 import json
-import chromadb
 
 # database
 cxn = sqlite3.connect('./data/mbr.db')
@@ -30,82 +29,81 @@ cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
 
 """# import datasets"""
 
-# dataset = load_dataset("gfhayworth/hack_policy", split='train')
-# mypassages = list(dataset.to_pandas()['psg'])
-
-# dataset_embed = load_dataset("gfhayworth/hack_policy_embed", split='train')
-# dataset_embed_pd = dataset_embed.to_pandas()
-# mycorpus_embeddings = torch_tensor(dataset_embed_pd.values)
-###########################################################################################################################
-"""# set up vector db"""
-from chromadb.config import Settings
-
-chroma_client = chromadb.Client(settings=Settings(
-    chroma_db_impl="duckdb+parquet",
-    persist_directory="./data/mychromadb/" # Optional, defaults to .chromadb/ in the current directory
-))
-collection = chroma_client.get_collection(name="benefit_collection")
-
-def vdb_rslt(qry,src,top_k=20):
-  results = collection.query(
-    query_embeddings=[ bi_encoder.encode(qry) ],
-    n_results=top_k,
-    where={"source": src},
-    )
-  return results
-##################################################################################################################################
-# Semantic Search Functions
-def rtrv(qry, src = 'H1036236000SB23.pdf', top_k=20):
-  rslts = vdb_rslt(qry,src, top_k) 
-  return rslts
-
-def rernk(query, collection=collection, top_k=20, top_n = 5):
-  rtrv_rslts = rtrv(query, top_k=top_k)
-  rtrv_ids = rtrv_rslts.get('ids')[0]
-  rtrv_docs = rtrv_rslts.get('documents')[0]
-
-  ##### Re-Ranking #####
-  cross_inp = [[query, doc] for doc in rtrv_docs]
-  cross_scores = cross_encoder.predict(cross_inp)
-
-  # Sort results by the cross-encoder scores
-  combined = list(zip(rtrv_ids, list(cross_scores)))
-  sorted_tuples = sorted(combined, key=lambda x: x[1], reverse=True)
-  sorted_ids = [t[0] for t in sorted_tuples[:top_n]]
-  predictions = collection.get(ids=sorted_ids, include=["documents","metadatas"])
-  return predictions
-
-def get_text_fmt(qry):
-  prediction_text = []
-  predictions = rernk(qry, collection=collection, top_k=20, top_n = 5)
-  docs = predictions['documents']
-  meta = predictions['metadatas']
-  for i in range(len(docs)):
-    result = Document(page_content=docs[i], metadata=meta[i]) 
-    prediction_text.append(result)
-  return prediction_text
+dataset = load_dataset("gfhayworth/hack_policy", split='train')
+mypassages = list(dataset.to_pandas()['psg'])
+
+dataset_embed = load_dataset("gfhayworth/hack_policy_embed", split='train')
+dataset_embed_pd = dataset_embed.to_pandas()
+mycorpus_embeddings = torch_tensor(dataset_embed_pd.values)
+
+def search(query, passages = mypassages, doc_embedding = mycorpus_embeddings, top_k=20, top_n = 1):
+    question_embedding = bi_encoder.encode(query, convert_to_tensor=True)
+    question_embedding = question_embedding #.cuda()
+    hits = util.semantic_search(question_embedding, doc_embedding, top_k=top_k)
+    hits = hits[0]  # Get the hits for the first query
+
+    ##### Re-Ranking #####
+    cross_inp = [[query, passages[hit['corpus_id']]] for hit in hits]
+    cross_scores = cross_encoder.predict(cross_inp)
+
+    # Sort results by the cross-encoder scores
+    for idx in range(len(cross_scores)):
+        hits[idx]['cross-score'] = cross_scores[idx]
+
+    hits = sorted(hits, key=lambda x: x['cross-score'], reverse=True)
+    predictions = hits[:top_n]
+    return predictions
+    # for hit in hits[0:3]:
+    #         print("\t{:.3f}\t{}".format(hit['cross-score'], mypassages[hit['corpus_id']].replace("\n", " ")))
+
+
+
+def get_text_fmt(qry, passages = mypassages, doc_embedding=mycorpus_embeddings):
+    predictions = search(qry, passages = passages, doc_embedding = doc_embedding, top_n=5, )
+    prediction_text = []
+    for hit in predictions:
+        page_content = passages[hit['corpus_id']]
+        metadata = {"source": hit['corpus_id']}
+        result = Document(page_content=page_content, metadata=metadata)
+        prediction_text.append(result)
+    return prediction_text
 
-##################################################################################################################################
 """# LLM based qa functions"""
 
-template = """You are a friendly AI assistant for the insurance company Humana. 
-Given the following extracted parts of a long document and a question, create a succinct final answer. 
+template = """You are a friendly AI assistant for the insurance company Humana. Given the following extracted parts of a long document and a question, create a succinct final answer. 
 If you don't know the answer, just say that you don't know. Don't try to make up an answer.
-If the question is not about Humana, politely inform the user that you are tuned to only answer questions about Humana.
+If the question is not about Humana, politely inform the user that you are tuned to only answer questions about Humana benefits.
 QUESTION: {question}
 =========
-{summaries}
+{context}
 =========
 FINAL ANSWER:"""
-PROMPT = PromptTemplate(template=template, input_variables=["summaries", "question"])
+PROMPT = PromptTemplate(template=template, input_variables=["context", "question"])
 
-chain_qa = load_qa_with_sources_chain(OpenAI(temperature=0), chain_type="stuff", prompt=PROMPT, verbose=True)
+chain_qa = load_qa_chain(OpenAI(temperature=0), chain_type="stuff", prompt=PROMPT)
+
+def get_text_fmt(qry, passages = mypassages, doc_embedding=mycorpus_embeddings):
+  predictions = search(qry, passages = passages, doc_embedding = doc_embedding, top_n=5, )
+  prediction_text = []
+  for hit in predictions:
+    page_content = passages[hit['corpus_id']]
+    metadata = {"source": hit['corpus_id']}
+    result = Document(page_content=page_content, metadata=metadata)
+    prediction_text.append(result)
+  return prediction_text
 
 def get_llm_response(message):
   mydocs = get_text_fmt(message)
-  responses = chain_qa({"input_documents":mydocs, "question":message})
+  responses = chain_qa.run(input_documents=mydocs, question=message)
   return responses
 
+# for x in xmpl_list:
+#   print(32*'=')
+#   print(x)
+#   print(32*'=')
+#   r = get_llm_response(x)
+#   print(r)
+
 """# Database query"""
 
 db = SQLDatabase.from_uri("sqlite:///./data/mbr.db")
@@ -115,10 +113,10 @@ llm = OpenAI(temperature=0)
 # model_name: str = "text-davinci-003"
 # instruction fine-tuned, sometimes referred to as GPT-3.5
 
-db_chain = SQLDatabaseChain(llm=llm, database=db, verbose=True, return_intermediate_steps=True)
+db_chain = SQLDatabaseChain(llm=llm, database=db, verbose=True)
 
 def db_qry(qry):
-  responses = db_chain('my mbr_id is 456 ;'+str(qry) ) ############### hardcode mbr id 456 for demo
+  responses = db_chain.run(query='my mbr_id is 456 ;'+str(qry) ) ############### hardcode mbr id 456 for demo
   return responses
 
 #db_qry('how many footcare visits have I had?')
@@ -180,38 +178,13 @@ def mrkl_rspnd(qry):
   response = mrkl({"input":str(qry) })
   return response
 
-def get_cot(r):
-  cot = '<p>'
-  try:
-    intermedObj = r['intermediate_steps']
-    cot +='<b>Input:</b> '+r['input']+'<br>'
-    for agnt_action, obs in intermedObj:
-      al = '<br>  '.join(agnt_action.log.split('\n') )
-      cot += '<b>AI chain of thought:</b> '+ al +'<br>'
-      if type(obs) is dict:
-        if obs.get('input_documents') is not None: 
-          for d in obs['input_documents']:
-            cot += '&nbsp;&nbsp;&nbsp;&nbsp;'+'<i>- '+str(d.page_content)+'</i>'+' <a href="'+ str(d.metadata['url']) +'">'+'''<span style="color: blue;">'''+str(d.metadata['page'])+'</span></a> '+'<br>'
-          cot += '<b>Observation:</b> '+str(obs['output_text']) +'<br><br>'
-        elif obs.get('intermediate_steps') is not None:
-          cot += '<b>Query:</b> '+str(obs.get('intermediate_steps')) +'<br><br>'
-        else:  
-          pass 
-      else:
-        cot += '<b>Observation:</b> '+str(obs) +'<br><br>'
-  except:
-    pass
-  cot += '</p>'  
-  return cot
-
 def chat(message, history):
     history = history or []
     message = message.lower()
     
     response = mrkl_rspnd(message)
-    cot = get_cot(response)
     history.append((message, response['output']))
-    return history, history, cot
+    return history, history
 
 
 

requirements.txt

@@ -1,5 +1,5 @@
 sentence-transformers
-# datasets
+datasets
 openai==0.26.1
 gradio
 # google-search-results
@@ -9,4 +9,3 @@ langchain
 requests==2.28.2
 git+https://github.com/openai/whisper.git
 boto3
-chromadb