Potpie Assignment

Table Of Contents

• Setup Instructions
• Technologies Used
• Workflow
• API Documentation
  • Process Document
  • Start Chat
  • Send Message
  • Chat History
• Possible Improvements

Setup Instructions

git clone git@github.com:dawkrish/potpie-test.git

cd potpie-test

python3 -m venv venv

pip install -r requirements.txt

Technologies used

• Python (>=3.10) because we use pattern-matching
• Flask for REST API
• ChromaDB as vector DB to store the embeddings (we have done persistently)
• PyPDF2 to get text from pdf files
• python-docx to get text from docx files
• dotenv to load environment variables
• HuggingFace Embeddings (to embed text and later used to make vectorstore)
• Langchain ties up everything
• llm used is ChatMistralAI, currently I have used all the tokens :"

Workflow

• User uploads the file.

• User creats an asset_id with the end point POST /api/document/process

  • text is extracted out for different extension types using get_file_text (embeddings.py)
  • save_embeddings (embeddings.py) is called to save the embedding.
    • creates a uuid and assigns to asset_id
    • get_huggingface_embeddings (embeddings.py) is used for embeddings (we have a default-embedder too get_default_embeddings)
    • chroma_collection.add to save the embedding
    • returns asset_id
  • the end point returns asset_id

• User creates a "thread_id" with the end point POST /api/chat/start

  • This end point takes the asset_id from the previous end point
  • It returns a thread_id .
  • The thread thread_to_asset_map exists to keep mappings, its a in-memory global variable
  • Now after getting the thread_id user can start chatting.

• End point POST /api/chat/message is used for to chat.

  • user_message the user prompt/query and thread_id which we got from last end point is needed
  • we get response from response_to_message (rag.py)
  • we use mistral-ai for now, it can be changed
  • The response will be in stream; Streaming can be turned off.
  • at this end point, the "chat-message" gets stored in the variable chat_history, another in-memory global variable
  • The chats are non-persistent. We can integrate sqlite3 to save the chats.

• End point GET /api/chat/history/<thread_id> returns the specific chat.

API Documentation

Generate from Mistral-AI

This API provides endpoints for processing documents and managing chat sessions. Below are the available endpoints and their usage.

Endpoints

1. Process Document

Endpoint: /api/document/process

Method: POST

Description: Processes a document and saves its embeddings.

Request Body:

{
    "file_path": "path/to/your/file"
}

Response:

• Success (200):

  {
      "asset_id": "generated_asset_id"
  }

• Error (400):

  {
      "error": "file_path is required"
  }

2. Start Chat

Endpoint: /api/chat/start

Method: POST

Description: Starts a new chat session based on a given asset ID.

Request Body:

{
    "asset_id": "your_asset_id"
}

Response:

• Success (200):

  {
      "thread_id": "generated_thread_id"
  }

• Error (400):

  {
      "error": "asset_id is required"
  }

• Error (404):

  {
      "error": "Invalid asset_id, no embeddings exist for it"
  }

3. Send Message

Endpoint: /api/chat/message

Method: POST

Description: Sends a user message and receives an AI response.

Request Body:

{
    "thread_id": "your_thread_id",
    "user_message": "your_message"
}

Response:

• Success (200):

  {
      "ai_response": "generated_ai_response"
  }

• Error (400):

  {
      "error": "thread_id is required"
  }

• Error (400):

  {
      "error": "user_message is required"
  }

• Error (404):

  {
      "error": "Invalid chat-thread id, create a new one"
  }

4. Get Chat History

Endpoint: /api/chat/history/<thread_id>

Method: GET

Description: Retrieves the chat history for a given thread ID.

Response:

• Success (200):

  {
      "chat_history": [
          {
              "timestamp": "2023-10-01T12:00:00Z",
              "user_message": "your_message",
              "ai_response": "generated_ai_response"
          }
      ]
  }

• Error (400):

  {
      "error": "thread_id does not exist in memory"
  }

Possible Improvements

• Better handling of chat-history, possibly using lang-graph
• Options to chose LLM
• Options for streaming
• Making code more flexible
• Python files documentation using Mistral-AI model Codestral (it is awesome.)