DMOM Chatbot

Official Code for the Paper VPHQA_ VIETNAMESE PREGNANCY HEALTH QUESTION ANSWERING DATASET

This project builds a Retrieval Augmented Generation (RAG) pipeline over the tungedng2710/Dmom_dataset dataset using a local vector database (Chroma) and an Ollama-served LLM.

Demo

Key choices:

• Vector DB: Chroma (local persistent store)
• Embeddings: via Ollama /api/embeddings (default: bge-m3:latest)
• Generator: selectable — Ollama /api/chat, Google Gemini REST, or Cerebras Cloud
• Dataset: tungedng2710/Dmom_dataset (fetched with datasets)

Prerequisites

• Python 3.9+
• pip
• Ollama installed and running, reachable at http://localhost:7860

Install Ollama

• Linux:
  • curl -fsSL https://ollama.com/install.sh | sh
  • Start on port 7860: export OLLAMA_HOST=0.0.0.0:7860 && ollama serve
• macOS:
  • brew install ollama (or download the app from ollama.com)
  • Start on port 7860: export OLLAMA_HOST=0.0.0.0:7860 && ollama serve
• Windows:
  • Install the Ollama app from ollama.com or via winget install Ollama.Ollama
  • Run it once, then in a terminal: setx OLLAMA_HOST 0.0.0.0:7860 and restart Ollama (or use WSL with the Linux command above).
• Docker (optional):
  • docker run -d --name ollama -p 7860:11434 -v ollama:/root/.ollama ollama/ollama
  • The API is then available at http://localhost:7860 (mapped to container’s 11434).

Verify Ollama

• curl http://localhost:7860/api/tags should return JSON (empty if no models yet).

Pull required models

• Generation (Ollama): ollama pull gpt-oss:20b
• Embeddings (default bge-m3:latest):
  • ollama pull bge-m3:latest

Install

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Configure (optional)

• Copy .env.example to .env and adjust values.
• Defaults:
  • OLLAMA_BASE_URL=http://localhost:7860
  • GENERATION_MODEL=gpt-oss:20b
  • EMBEDDING_MODEL=bge-m3:latest
  • CHROMA_DIR=./data/chroma
  • CHAT_BACKEND=ollama (set to gemini or cerebras to switch cloud providers)
  • For Gemini: set GEMINI_API_KEY and optionally GEMINI_MODEL (e.g., gemini-1.5-flash)
  • For Cerebras: set CEREBRAS_API_KEY and optionally CEREBRAS_MODEL (e.g., llama-4-scout-17b-16e-instruct)

Commands

1. Ingest from local CSV (recommended for dmom_data.csv)

python -m tonrag.cli ingest \
  --csv data/dmom_data.csv \
  --text-field Reference \
  --id-field no \
  --chunk-size 800 \
  --chunk-overlap 120

2. Ingest from Hugging Face dataset

python -m tonrag.cli ingest \
  --dataset tungedng2710/Dmom_dataset \
  --split train \
  --text-field context \
  --id-field id \
  --chunk-size 800 \
  --chunk-overlap 120

Notes:

• If you are unsure of field names, run:
  • CSV: python -m tonrag.cli inspect --csv data/dmom_data.csv
  • HF: python -m tonrag.cli inspect --dataset tungedng2710/Dmom_dataset --split train
• The script tries a few sensible defaults and will suggest possible column names.

3. Answer a single question

python -m tonrag.cli query \
  --question "<your question>" \
  --top-k 5

Use Gemini or Cerebras instead of Ollama for this run:

python -m tonrag.cli query --question "<your question>" --llm gemini
python -m tonrag.cli query --question "<your question>" --llm cerebras

4. Evaluate on the dataset (quick lexical match)

# CSV example
python -m tonrag.cli eval \
  --csv data/dmom_data.csv \
  --question-field instruction \
  --answer-field output \
  --top-k 5 \
  --limit 50

# HF example
python -m tonrag.cli eval \
  --dataset tungedng2710/Dmom_dataset \
  --split validation \
  --question-field question \
  --answer-field answer \
  --top-k 5 \
  --limit 50

Project Structure

• tonrag/config.py – environment/config defaults
• tonrag/embeddings.py – Ollama embeddings client; ST fallback if available
• tonrag/llm.py – Ollama chat client (non-streaming)
• tonrag/vectorstore.py – Chroma wrapper
• tonrag/dataset.py – dataset utilities and column auto-detection
• tonrag/chunking.py – simple text chunker
• tonrag/rag.py – retrieval + prompt assembly + generation
• tonrag/cli.py – CLI entry points (ingest/query/eval/inspect)
• app/server.py – minimal web app (stdlib) serving app/static/ and /api/chat
• app/static/ – frontend assets (index.html, style.css, app.js)

Troubleshooting

• If embeddings fail, ensure the embedding model is pulled and available in Ollama: ollama pull bge-m3:latest.
• If generation fails, ensure gpt-oss:20b is available: ollama pull gpt-oss:20b.
• If dataset download fails, ensure your environment has internet access and the datasets package can reach Hugging Face.

Web App

• FastAPI server (recommended): uvicorn app.main:app --host 0.0.0.0 --port 7865
• Open: http://localhost:7865
• Endpoints:
  • GET / – serves UI
  • POST /api/chat – body: { "message": "...", "top_k": 5, "llm": "ollama|gemini|cerebras", "llm_api_key": "optional" }
  • GET /health
• The app uses the same RAG pipeline and Chroma store.

Legacy stdlib server (optional): python app/server.py --port 7865

Chat UI

• Clean bubble layout with avatars, typing indicator, and copy buttons.
• Collapsible “Contexts” under each assistant message to inspect retrieved sources.
• Shift+Enter for newline, Enter to send, clear chat, adjustable Top‑K.
• Answers render Markdown (headings, lists, links, code blocks) safely in the UI.
• CLI: add --strip-markdown to print plain‑text answers.
• LLM selector in the header to switch between Ollama, Gemini, or Cerebras per request.

License

• For this template, no explicit license is added; adapt as needed.