CodeWare RAG Chatbot (Python Backend Engineer AI/RAG Focus)

A modular Q&A chatbot API that combines:

• RAG pipeline (LangChain + ChromaDB + Ollama local LLM)
• Rule-based flow trigger (from codeware_bot_flow.json)

Supports Bangla, English, and Banglish queries.

---

Features

• Rule-based flow: Triggers external API for service flows (Packages, Bill Pay, New Connection, etc.)
• RAG pipeline: Semantic search over company docs, answer generation via local LLM
• Modular code: SOLID, DRY, easy to extend

---

Setup Instructions

1. Prerequisites

• Python 3.10+
• Ollama installed and running (for local LLM)
• Conda or venv for Python environment
• (Recommended) Mac M1: use a small/quantized Ollama model (e.g., phi3:Mini)
• This codebase uses phi3:Mini. If u wish to use any other models, you can just change it in config.py

2. Clone & Environment

# Clone the repo
# git clone <your-repo-url>
cd directory

# Create and activate environment
conda create -n chatbot-rag python=3.10 -y
conda activate chatbot-rag

# Install dependencies (conda)
conda install --file requirements.txt
# (Optional, to silence deprecation warnings, use pip for these two only if needed)
# pip install -U langchain-chroma langchain-ollama

3. Prepare Data

• Place your company info .txt files in the company_info/ folder.
• Ensure codeware_bot_flow.json is present in the project root.

4. Ingest Data (Embedding)

python embed_company_info.py

• This will chunk, embed, and store your docs in ChromaDB (chromadb_store/).

5. Start Ollama (if not running)

ollama serve &
ollama pull phi3:mini  # or another small model

6. Run the API Server

uvicorn app:app --reload

---

API Endpoints

POST /chat

Input:

{
  "user_id": "12345",
  "question": "I want to know about packages"
}

Logic:

1. Checks for rule-based triggers (Packages, Bill Pay, etc.)
   • If matched, calls /chatbot and returns trigger info.
2. Otherwise, runs RAG pipeline (semantic search + LLM answer).

Response (RAG):

{
  "triggered": false,
  "answer": "Our packages include Basic, Premium, and Business plans...",
  "sources": ["doc1.txt", "doc2.txt"]
}

Response (Rule-based):

{
  "triggered": true,
  "trigger_id": "...",
  "matched_keyword": "Packages",
  "answer": "Rule-based flow triggered for keyword 'Packages'. (Simulated external API call.)"
}

POST /chatbot

Input:

{
  "user_id": "12345",
  "trigger_id": "..."
}

Response:

{
  "user_id": "12345",
  "trigger_id": "...",
  "message": "Simulated external API response for trigger_id ..."
}

---

Example Usage

Test with curl:

curl -X POST http://127.0.0.1:8000/chat \
  -H "Content-Type: application/json" \
  -d '{"user_id":"u1","question":"What packages do you offer?"}'

---

Notes

• For best performance on Mac M1, use a quantized/small Ollama model.
• If you see deprecation warnings, install the new adapters as above.
• To extend: add new flows to codeware_bot_flow.json or new docs to company_info/ and re-run embedding.

---

Submission

• Push code to a GitHub repo.
• Include this README.
• Ensure all endpoints and logic match the project spec.

Postman Example

Below are some Postman examples for interacting with the chatbot:

1. 
2. 
3. 

Discussion

English works fine for both trigger-flow-based and LLM based Q&A. To enable bangla Language support I alternate the embedding model from all-MiniLM-L6-v2 to csebuetnlp/banglishbert. But still the accuracy did not increase. I tried to increase chunk size and relevent info grouping for better chunking. Though it seem to have a good effect but not quite. I have found the problem. The string match for Bangla language not working. I think resolving this will ensure the Bangla Q&A.