Upload 4 files

app.py

@@ -4,8 +4,7 @@ import os
 import gradio as gr
 import langchain
 import pickle
-from langchain.vectorstores import Weaviate
-from langchain import OpenAI
+from langchain.llms import HuggingFaceHub
 
 from chain import get_new_chain1
 
@@ -15,23 +14,24 @@ def get_faiss_store():
         return faiss_store
 
 
-def set_openai_api_key(api_key, agent):
-    if api_key:
-        os.environ["OPENAI_API_KEY"] = api_key
-        vectorstore = get_faiss_store()
+def load_model():
 
-        rephraser_llm = OpenAI(model_name="text-davinci-003", temperature=0)
-        final_output_llm = OpenAI(model_name="text-davinci-003", temperature=0, max_tokens=-1)
+    print(langchain.__file__)
 
-        qa_chain = get_new_chain1(vectorstore, rephraser_llm, final_output_llm)
-        os.environ["OPENAI_API_KEY"] = ""
-        return qa_chain
+    vectorstore = get_faiss_store()
+
+    flan_ul = HuggingFaceHub(repo_id="google/flan-ul2", 
+                                model_kwargs={"temperature":0.1, "max_new_tokens":200},
+                                huggingfacehub_api_token="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")
+
+    qa_chain = get_new_chain1(vectorstore, flan_ul, flan_ul, isFlan=True)
+    return qa_chain
 
 
 def chat(inp, history, agent):
     history = history or []
     if agent is None:
-        history.append((inp, "Please paste your OpenAI key to use"))
+        history.append((inp, "Please click Load Model or wait for model to load"))
         return history, history
     print("\n==== date/time: " + str(datetime.datetime.now()) + " ====")
     print("inp: " + inp)
@@ -49,12 +49,10 @@ with block:
     with gr.Row():
         gr.Markdown("<h3><center>Hugging Face Doc Search</center></h3><p>Ask questions about the Hugging Face Transformers Library</p>")
 
-        openai_api_key_textbox = gr.Textbox(
-            placeholder="Paste your OpenAI API key (sk-...)",
-            show_label=False,
-            lines=1,
-            type="password",
-        )
+        load_model_button = gr.Button(
+            value="Load Model",
+            variant="secondary"
+        ).style(full_width=False)
 
     chatbot = gr.Chatbot()
 
@@ -90,10 +88,6 @@ with block:
     submit.click(chat, inputs=[message, state, agent_state], outputs=[chatbot, state])
     message.submit(chat, inputs=[message, state, agent_state], outputs=[chatbot, state])
 
-    openai_api_key_textbox.change(
-        set_openai_api_key,
-        inputs=[openai_api_key_textbox, agent_state],
-        outputs=[agent_state],
-    )
+    load_model_button.click(load_model, outputs=[agent_state])
 
 block.launch(debug=True)

chain.py

@@ -4,14 +4,12 @@ import pathlib
 import pickle
 from typing import Dict, List, Tuple
 
-import weaviate
-from langchain import OpenAI, PromptTemplate
+from langchain import PromptTemplate
 from langchain.chains import LLMChain
 from langchain.chains.base import Chain
 from langchain.chains.combine_documents.base import BaseCombineDocumentsChain
 from langchain.chains.conversation.memory import ConversationBufferMemory
 from langchain.chains.question_answering import load_qa_chain
-from langchain.embeddings import OpenAIEmbeddings
 from langchain.prompts import FewShotPromptTemplate, PromptTemplate
 from langchain.prompts.example_selector import \
     SemanticSimilarityExampleSelector
@@ -42,15 +40,23 @@ class CustomChain(Chain, BaseModel):
         else:
             new_question = question
         print(new_question)
-        docs = self.vstore.similarity_search(new_question, k=4)
+        docs = self.vstore.similarity_search(new_question, k=3)
         new_inputs = inputs.copy()
         new_inputs["question"] = new_question
         new_inputs["chat_history"] = chat_history_str
         answer, _ = self.chain.combine_docs(docs, **new_inputs)
-        return {"answer": answer}
 
+        ## Dedupe source list
+        source_list = [doc.metadata['source'] for doc in docs]
 
-def get_new_chain1(vectorstore, rephraser_llm, final_output_llm) -> Chain:
+        source_string = "\n\n*Sources:* "
+        for i, source in enumerate(set(source_list)):
+            source_string += f"[[{i}](https://{source})]"
+
+        final_answer = answer + source_string
+        return {"answer": final_answer}
+
+def get_new_chain1(vectorstore, rephraser_llm, final_output_llm, isFlan) -> Chain:
     _eg_template = """## Example:
 
     Chat History:
@@ -73,7 +79,7 @@ def get_new_chain1(vectorstore, rephraser_llm, final_output_llm) -> Chain:
     #### LOAD VSTORE WITH REPHRASE EXAMPLES
     with open("rephrase_eg.pkl", 'rb') as f:
         rephrase_example_selector = pickle.load(f)
-    
+
     prompt = FewShotPromptTemplate(
         prefix=_prefix,
         suffix=_suffix,
@@ -89,25 +95,38 @@ def get_new_chain1(vectorstore, rephraser_llm, final_output_llm) -> Chain:
         input_variables=["page_content", "source"],
     )
 
-    template = """You are an AI assistant for the open source transformers library provided by Hugging Face. The documentation is located at https://huggingface.co/docs/transformers.
-You are given the following extracted parts of a long document and a question. Provide a conversational answer with a hyperlink to the documentation. Do NOT add .html to the end of links. Make sure to bold link text.
-You should only use hyperlinks that are explicitly listed as a source in the context. Do NOT make up a hyperlink that is not listed.
-If the question includes a request for code, provide a code block directly from the documentation.
+    gpt_template = """You are an AI assistant for the open source transformers library provided by Hugging Face. The documentation is located at https://huggingface.co/docs/transformers.
+- You are given extracted parts of a long document and a question. 
+- Provide a conversational answer with a hyperlink to the documentation based on the "source". 
+- Do NOT add .html to the end of links. Make sure to bold link text.
+- You should only use hyperlinks that are explicitly listed as a source in the context. Do NOT make up a hyperlink that is not listed.
+- If the question includes a request for code, provide a code block directly from the documentation.
+- If you don't know the answer, just say "Hmm, I'm not sure." Don't try to make up an answer.
+- If the question is not about Hugging Face Transformers, politely inform them that you are tuned to only answer questions about Transformers.
+
 For example, if someone asks how to install Transformers, you should say:
 
-You can install with pip, for more info view the **(documentation)**[https://huggingface.co/docs/transformers/installation]
+You can install with pip: 
 '''py
 pip install transformers
 '''
+**(Source)**[https://huggingface.co/docs/transformers/main/en/installation]
 
-If you don't know the answer, just say "Hmm, I'm not sure." Don't try to make up an answer.
-If the question is not about Hugging Face Transformers, politely inform them that you are tuned to only answer questions about Transformers.
 Question: {question}
 =========
 {context}
 =========
 Answer in Markdown:"""
-    PROMPT = PromptTemplate(template=template, input_variables=["question", "context"])
+
+    flan_template = """
+    {context}
+    Based on the above documentation, answer the user's question in markdown: {question}"""
+
+    PROMPT = PromptTemplate(template=gpt_template, input_variables=["question", "context"])
+
+    if isFlan:
+        PROMPT = PromptTemplate(template=flan_template, input_variables=["question", "context"])
+
     doc_chain = load_qa_chain(
         final_output_llm,
         chain_type="stuff",
@@ -120,8 +139,10 @@ Answer in Markdown:"""
 
 def _get_chat_history(chat_history: List[Tuple[str, str]]):
     buffer = ""
-    for human_s, ai_s in chat_history:
+    for human_s, ai_s in chat_history[-2:]:
         human = f"Human: " + human_s
         ai = f"Assistant: " + ai_s
         buffer += "\n" + "\n".join([human, ai])
+
+    
     return buffer

ingest.py

@@ -4,9 +4,10 @@ from pathlib import Path
 from markdown import markdown
 
 import pickle
+import re
 from bs4 import BeautifulSoup
 from langchain.text_splitter import CharacterTextSplitter
-from langchain.embeddings import HuggingFaceInstructEmbeddings, OpenAIEmbeddings
+from langchain.embeddings import HuggingFaceInstructEmbeddings
 from langchain.vectorstores import FAISS
 from InstructorEmbedding import INSTRUCTOR
 
@@ -16,7 +17,9 @@ def clean_data(data):
     html = markdown(data)
     soup = BeautifulSoup(html, "html.parser")
     text = ''.join(soup.findAll(text=True))
-    return "\n".join([t for t in text.split("\n") if t])
+    cleaned_text = re.sub(r"<!--.*?-->", "", text, flags=re.DOTALL)
+    print(cleaned_text)
+    return "\n".join([t for t in cleaned_text.split("\n") if t])
 
 docs = []
 metadatas = []