Lots of people using LangChain are struggling with working with multiple functions. One common issue is the ability to create a simple chat bot that uses the following:
- Prompt Template
- Chat History
- Vector DB (Pinecone)
It uses ConversationalRetrievalChain for the retrieval of the DB.
It uses my Prompt Template (template), the context from the embedding search and the user's question. Provides an answer. Then, in the "background", for follow up question, it uses the follow up template (_template) to combine the new question with the previous answer, to provide an additional context. That additional context is not sent to chatGPT, all it does is create a new question on its own, which would then be passed to the query, retrieve full context from docs and finally use the query/full context to send to chatGPT.
Example Process: KB about house painting
- User Question 1: "How do I paint a door?" -> Vector/Embeddings search -> Get full context -> Send question with context -> Return Answer 1: "Take a brush out, open can of paint, mask the door frame, etc."
- User Question 2: "Do I need a tool to open it?" -> Background process: Combine Answer 1 with question 2 and create question 2-A: "When painting, what tools can be used to open the can of paint?"
- User Question 2 becomes Question 2-A -> Vector/Embeddings search -> Get full context -> Send question with context -> Return Answer 2: "Just use a screwdriver!"
This may get a little "buggy" with lots of follow up questions but should give a base to use for better coding. Also, it needs a lot of finesse with the prompt engineering based on your context and expected tone of answers, so will need to try variations there to be flexible enough for answers.
- This code is only to help to create your own. It is far from perfect. No support is provided.
- If you find it useful and have comments for suggestions or improvements, please help the community and open an issue in this repository so that it can be shared with others.
- If you add features to it, please also share.
- If you have code to share based on wished improvements below, please also add to an issue.
- As time permit, I will try to update code with any suggestions for improvements.
- Token count was mentioned by someone
- Ability to cache some queries
- Find a way to only add vectors to DB if docs have changed or been added