joebordes/doc_chat_api

Create a production level scalable chat bot to respond from the fed data

Production level scalable Documents ChatBot

About

ChatBot to do conversation based on stored data provided by the user. Here is flow diagram

Here are few advantages.

• The chatBot uses a Retriever-Generator base module to reduce costs. The Retriever fetches the text of concern while the Generator creates a response from the fetched content.
• OpenAI GPT3.5, and open-source models are supported
• Embeddings are created and stored in a Milvus vector database.
• History is stored in SQLite

Prerequisite

• Install docker engine

• Install docker compose

• Install and run milvus. See bottom of page for more info.

• Install langchain / LlamaIndex

• Download open-source model weights from GPT4All. The models I have tested is

  • ggml-gpt4all-j.bin (commercial licensable)
  • ggml-gpt4all-l13b-snoozy.bin (non-commercial licensable)

• Put openAI API key in example.env in case if you want to use openAI model and replace example.env to .env

API Documentation

This documentation provides information about the API endpoints available in the FastAPI-based API.

Models

DocModel

Represents the model for adding documents for ingestion.

Field	Type	Description
dir_path	str	The directory path of the documents to ingest.
embeddings_name	str (optional)	The name of the embeddings ['openai', 'sentence'] (default: 'openai').
collection_name	str (optional)	The name of the collection (default: 'LangChainCollection').
drop_existing_embeddings	bool (optional)	Whether to drop existing embeddings (default: False).

QueryModel

Represents the model for processing user queries.

Field	Type	Description
text	str	The text for the query.
session_id	uuid4	The session ID for the query.
llm_name	str (optional)	The name of the language model ['openai', 'llamacpp', 'gpt4all'] (default: 'openai').
collection_name	str (optional)	The name of the collection (default: 'LangChainCollection').

DeleteSession

Represents the model for deleting a session from the database.

Field	Type	Description
session_id	uuid4	The session ID to delete.

Endpoints

POST /doc_ingestion

Endpoint to add documents for ingestion.

Request

• Body Parameters:
  • doc (DocModel): The document ingestion details.

Response

• Status Code: 200 (OK)
• Body: {"message": "Documents added successfully"}

POST /query

Endpoint to process user queries.

Request

• Body Parameters:
  • query (QueryModel): The user query details.

Response

• Body: {"answer": str, "cost": dict, "source":list}

-- answer : answer from the documents
-- cost "cost": {
    "successful_requests": int,
    "total_cost": float,
    "total_tokens": int,
    "prompt_tokens": int,
    "completion_tokens": int
  },
  --source: list of str showing source of extracted answer

POST /delete

Endpoint to delete a session from the database.

Request

• Body Parameters:
  • session (DeleteSession): The session deletion details.

Response

• Body: The response message indicating the success or failure of the deletion operation.

Example Usage

Adding Documents for Ingestion

$ curl -X POST -H "Content-Type: application/json" -d '{
    "dir_path": "/path/to/documents"
}' http://localhost:8000/doc_ingestion

Processing User Queries

$ curl -X POST -H "Content-Type: application/json" -d '{
    "text": "User query",
    "session_id": "9c17659b-f3f6-45c5-8590-1a349102512b"
}' http://localhost:8000/query

Deleting a Session

$ curl -X POST -H "Content-Type: application/json" -d '{
    "session_id": "9c17659b-f3f6-45c5-8590-1a349102512b"
}' http://localhost:8000/delete

TODOs

• Change pre-defined prompt
• Filter data (profanity/offensive language)
• Allow open-source LLMs
• Streaming response
• Make memory optional to speedup response.
• Add docker/docker compose

Guide to run

docker compose -f docker-compose.milvus.yml up -d
docker compose -f docker-compose.app.yml up -d

Note:

The Chatbot is also implemented using haystack