Sinhala Hybrid RAG Chatbot (Offline)

Complete Sinhala chatbot system using Hybrid RAG (JSON + Text), Streamlit UI, FAISS retrieval, and Ollama local inference.

Assignment Coverage

• Fully offline capable at runtime
• Ollama local LLM inference (http://localhost:11434/api/generate)
• Streamlit chat interface
• Sinhala Unicode input/output
• In-session memory for recent conversation turns (last 10)
• Hybrid retrieval flow:
  1. JSON semantic retrieval (Top 2)
  2. Text semantic retrieval from selected topics (Top 3)
  3. Context merge and grounded generation

Project Structure

project/
├── app.py
├── chatbot/
│   ├── hybrid_retriever.py
│   ├── json_retriever.py
│   ├── text_retriever.py
│   ├── embeddings.py
│   ├── prompt.py
│   ├── ollama.py
│   ├── memory.py
│   ├── build_indexes.py
│   └── test_queries.py
├── data/
│   ├── knowledge.json
│   └── documents/
│       ├── headache.txt
│       ├── stress.txt
│       └── ...
├── vectorstore/
│   ├── json_index.faiss
│   └── text_index.faiss
├── requirements.txt
└── README.md

Offline Requirements

1. Ollama must be installed locally.
2. Gemma model must already exist locally.
3. Sentence-transformers embedding model must be locally available.

Runtime code uses local-only embedding loading by default (EMBEDDING_LOCAL_ONLY=1).

Install

pip install -r requirements.txt

Run Ollama Offline

ollama serve
ollama pull gemma

After model pull is completed once, usage is local/offline.

Build FAISS Indexes

python -m chatbot.build_indexes

This generates:

• vectorstore/json_index.faiss
• vectorstore/text_index.faiss

Run Streamlit Chatbot

streamlit run app.py --server.port 8501

Open:

• http://localhost:8501

Prompt Policy

The exact assignment prompt is used in chatbot/prompt.py with strict fallback:

"මට ඒ පිළිබඳ ප්‍රමාණවත් තොරතුරු නොමැත"

Testing (20 Sinhala Queries)

Run retrieval tests:

python -m chatbot.test_queries

The script includes 20 Sinhala test cases and checks retrieval correctness by expected topic.

Notes

• No keyword matching is used for retrieval.
• Full dataset is never sent to the model.
• Only retrieved context is sent to Ollama.