A retrieval-augmented chatbot that helps people explore and debate the writings of Dr. B. R. Ambedkar in calm, plain English. The agent grounds every answer in the PDF corpus stored in Ambedkar_Writings/, and keeps the tone empathetic, mature, and constructive even when faced with disagreement.

• Vectorise Ambedkar's collected writings with OpenAI embeddings and an Annoy index stored locally in data/
• Fast retrieval of relevant passages for each question, with volume and page level references
• Persona-aware chat pipeline that encourages reflective, respectful dialogue around points of disagreement
• Typer-powered CLI for ingestion, status checks, and interactive conversations
• Poetry-managed environment with reproducible dependency graph

You can drive the project either with uv or with Poetry. Pick the flow that matches your toolchain.

1. Create and activate the virtual environment

   cd /Users/sundarraghavanl/PycharmProjects/ambedkar_chatbot
   uv venv --python /usr/bin/python3 .venv
   source .venv/bin/activate

2. Install the project in editable mode

   uv pip install --upgrade pip setuptools wheel
   uv pip install -e .

3. Create a .env file (or export the variables in your shell):

   cat <<'ENV' > .env
   OPENAI_API_KEY=sk-your-key
   # Optional overrides
   # EMBED_MODEL=text-embedding-3-small
   # CHAT_MODEL=gpt-4o-mini
   # CHUNK_SIZE=320
   # CHUNK_OVERLAP=60
   # TOP_K=6
   ENV

4. Build the vector store (downloads embeddings for all PDFs and writes Annoy index + metadata):

   ambedkar-chatbot ingest

   The first run can take a while and incurs OpenAI embedding costs. Artefacts are written to data/.
5. Chat with the companion:

   ambedkar-chatbot chat

   Type exit (or press Ctrl+D) to leave the conversation.

1. Install dependencies

   poetry install

2. Create a .env file (or export the variables in your shell):

   cat <<'ENV' > .env
   OPENAI_API_KEY=sk-your-key
   # Optional overrides
   # EMBED_MODEL=text-embedding-3-small
   # CHAT_MODEL=gpt-4o-mini
   # CHUNK_SIZE=320
   # CHUNK_OVERLAP=60
   # TOP_K=6
   ENV

3. Build the vector store:

   poetry run ambedkar-chatbot ingest

4. Chat with the companion:

   poetry run ambedkar-chatbot chat

   The CLI behaviour matches the uv commands above.

• ambedkar-chatbot ingest --incremental — append only new material; defaults to full rebuild
• ambedkar-chatbot info — quick health check for index/metadata files
• ambedkar-chatbot chat --top-k 8 — override the number of context chunks

The chatbot blends the warmth of a community educator with Ambedkar's analytical rigor:

• Uses plain, inclusive English and avoids jargon when possible
• Surfaces relevant citations (volume + page) whenever passages inform the reply
• Engages disagreements patiently, validating the other person's curiosity before clarifying Ambedkar's stance
• Encourages further reading, offering actionable suggestions to continue learning

Ambedkar_Writings/    # Source PDFs (input corpus)
data/                 # Generated embeddings, metadata, and Annoy index (git-ignored)
src/ambedkar_chatbot/ # Python package (config, ingest pipeline, vector store, chatbot, CLI)
poetry.toml / .venv   # Poetry config and in-project virtualenv (ignored)

• Add automated tests that mock OpenAI responses for repeatable CI runs
• Expose the persona via a small FastAPI or Streamlit frontend
• Add incremental ingestion heuristics (hashing chunks, skipping unchanged PDFs)
• Experiment with local embedding models when GPU resources are available

Released under the MIT License.