Ask the Pali Canon

Live: https://pcaisearch.netlify.app

Semantic search and AI-synthesized answers over the Pali Canon (DN, MN, AN, SN, Dhammapada, Itivuttaka, Udāna, Sutta Nipāta, Theragāthā, Therīgāthā, Khuddakapāṭha), grounded in Thanissaro Bhikkhu's English translations from dhammatalks.org.

Features

• Semantic search — multilingual embeddings (paraphrase-multilingual-MiniLM-L12-v2) retrieve relevant verses in English or Pali
• Hybrid retrieval — dense vector search fused with BM25 sparse retrieval via Reciprocal Rank Fusion; canonical suttas matched by title are surfaced even when vector search misses
• Query expansion — LLM generates alternative phrasings to improve recall
• Cross-encoder reranking — results reordered by relevance before display
• AI Synthesis — LLM answers your question using only retrieved context, with inline citations ([DN 1:1], [SN 46.20:14])
• Citation guardrail — distinguishes true hallucinations (non-existent sutta) from canonical misses (real sutta not in retrieved context)
• Nikaya filter — filter search and synthesis by collection (DN, MN, SN, AN, DHP, ITI, UD, STNP, THAG, THIG, KHP); click to switch, ⌘/Ctrl-click to combine
• Resume-capable indexing — indexing can be interrupted and resumed without re-embedding

Architecture

frontend/          Next.js 16 (App Router)
backend/           FastAPI + asyncio
  app/
    main.py        API endpoints (/synthesize, /stream), rate limiting
    core/
      indexing.py  SuttaParser, EmbeddingManager (fastembed / ONNX Runtime)
    services/
      search_pipeline.py   Query expansion → retrieval → reranking
      retriever.py         Dense vector retrieval (Qdrant)
      bm25_retriever.py    Sparse BM25 retrieval, fused via RRF
      sutta_title_index.py Sutta title BM25 boost
      fusion.py            Reciprocal Rank Fusion for hybrid retrieval
      guardrail.py         Citation verification (hallucination vs canonical miss)
      citation_oracle.py   Validates sutta IDs and verse numbers
      pali_dictionary.py   Pāḷi term → English passage hints for reranking
data/
  fetch_thanissaro.py  Download Thanissaro Bhikkhu epub from dhammatalks.org → local JSON
  process_dumps.py Embed & upsert into Qdrant
docs/adr/          Architecture decision records
tests/             pytest suites (backend)

Stack: FastAPI · Qdrant Cloud · fastembed (ONNX Runtime) · BM25 sparse retrieval · cross-encoder/ms-marco-MiniLM-L-6-v2 · Llama 3.1 8B for query expansion and synthesis (via NVIDIA API; synthesis model set via LLM_MODEL env var, defaults to meta/llama-3.3-70b-instruct) · Next.js · Tailwind CSS

Deployment

The frontend is on Netlify (pcaisearch.netlify.app). The backend runs on DigitalOcean App Platform (paid, ~$7–12/month), auto-deploying from the master branch on push. Vectors are stored in Qdrant Cloud (free tier). LLM calls go to the NVIDIA API (free tier). User feedback is stored durably in Supabase (free tier) — survives redeploys. The LLM_MODEL, SUPABASE_URL, and SUPABASE_KEY env vars are set in the App Platform dashboard.

Prerequisites

• Docker (for Qdrant)
• Python 3.10+
• Node.js 20+
• An NVIDIA API key

Setup

1. Start Qdrant

Option A — local Docker:

docker run -d -p 6333:6333 -v ~/qdrant_storage:/qdrant/storage qdrant/qdrant

Option B — Qdrant Cloud: create a free cluster at cloud.qdrant.io, then set QDRANT_URL and QDRANT_API_KEY in your environment before running the backend or indexing scripts.

2. Index the Pali Canon

# Download all 11 nikayas from the dhammatalks.org epub (Thanissaro Bhikkhu translations)
python3 data/fetch_thanissaro.py

# Embed and index into Qdrant (takes several minutes)
PYTHONPATH=. python3 data/process_dumps.py

Indexing can be interrupted and resumed — already-indexed suttas are skipped automatically.

3. Install frontend dependencies

cd frontend && npm install

4. Configure environment

cp frontend/.env.local.example frontend/.env.local
# Edit frontend/.env.local and set NEXT_PUBLIC_API_URL if needed

Running

Terminal 1 — Backend:

PYTHONPATH=. NVIDIA_API_KEY=your_key uvicorn backend.app.main:app --reload

Terminal 2 — Frontend:

cd frontend && npm run dev

Open http://localhost:3000.

API

Endpoint	Description
GET /synthesize?q=…&top_k=10	AI answer with citations, hallucinations, canonical_misses, and is_faithful flag
GET /stream?q=…&top_k=10&nikayas=DN	Streaming synthesis (SSE); same nikayas filter supported
POST /feedback	Submit thumbs-up/down feedback on a synthesis answer, with optional category and notes; stored in Supabase (production) or local SQLite (dev)

Rate limits: 10 req/min for synthesis and streaming.

Running Tests

PYTHONPATH=. python -m pytest tests/backend/ -q

License

CC BY-NC 4.0 — see LICENSE.

Sutta texts © Thanissaro Bhikkhu, sourced from dhammatalks.org.