🪔 InnerOracle

Life questions, answered only from the Bhagavad Gita & the Ashtavakra Gita — with every claim cited to the verse it came from.

→ Try it live at inneroracle.cc

---

What it is

InnerOracle is a grounded AI assistant for life questions. Instead of answering from a model's general knowledge, it retrieves relevant verses from two classical wisdom texts — the Bhagavad Gita and the Ashtavakra Gita — and answers strictly from them, citing each claim inline. When the two texts disagree, it shows both perspectives rather than blending them. When the texts don't cover something, it says so instead of guessing.

Ask anything — "How do I deal with anxiety about results?", "What is the self?", "How should I handle anger?" — and get a concise, sourced answer you can trace back to the original Sanskrit.

What makes it different

• 📖 Grounded, not guessed — answers come only from the retrieved verses. Every substantive claim carries an inline citation like [BG 2.47] or [AG 1.15].
• ✅ Self-checking — a separate model verifies each answer against the source verses and shows a "grounded in the texts" / "check sources" badge, so you know how trustworthy it is.
• 🔍 Tap to see the shloka — tap any citation to reveal the original Devanagari verse, transliteration, and translation.
• ⚖️ Honest contrast — the Gita's path of selfless action vs. the Ashtavakra Gita's non-dual "you were never the doer" — presented side by side. Focus on one text or compare both.
• 💬 Remembers the conversation — ask follow-ups like "tell me more" or "what about anger?" and it keeps context.
• 🔒 Private by design — conversations live only in your browser; nothing is stored on a server.

See it in action

A grounded answer — inline citations to each verse and a "grounded in the texts" verification badge. Light and dark themes:

Welcome screen — dark theme:

On mobile:

How it works

your question
   → (for follow-ups) rewritten into a standalone search query
      → hybrid retrieval: dense semantic + keyword search, fused and balanced across both texts
         → the model writes a concise answer grounded in those verses, streamed live
            → a verifier checks the answer against the verses → "grounded" / "check sources" badge

Sanskrit shlokas are kept separate from translations and commentary when the texts are ingested, then recombined into full verse context at answer time — so retrieval stays precise while answers stay readable.

Built with

Retrieval-Augmented Generation · Anthropic Claude (answers + verification) · FastAPI with live token streaming · Qdrant hybrid vector search (semantic + keyword, RRF fusion) · ONNX embeddings · SQLite · a dependency-light vanilla-JS web app · deployed with Docker behind a Cloudflare Tunnel.

Run it locally

git clone <this repo> && cd inneroracle
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env          # set ANTHROPIC_API_KEY
uvicorn app.main:app --reload # → http://localhost:8000

This assumes the knowledge base already exists under data/ (verses.db, qdrant_local/, bm25.json). If you're starting from scratch, build it first — see step 1 below.

Deploy your own

Prerequisites: a VPS with ≥ 4 GB RAM and Docker, a domain on Cloudflare, an Anthropic API key, and a workstation (your laptop) for the one-time ingestion/indexing.

1. Build the knowledge base on your workstation — never on a small VPS (indexing loads the embedding model and will OOM a 4 GB server):

   pip install -r requirements.txt -r requirements-ingest.txt -r requirements-embed.txt
   # one-time extraction: EPUB/PDF → structured verses in data/verses.db
   python -m ingest data/sources/<gita>.epub --book bhagavad_gita
   python -m ingest data/sources/<ashtavakra>.epub --book ashtavakra_gita
   # embed + index: builds data/qdrant_local/ (vectors) and data/bm25.json (keyword index)
   python -m ingest.index

2. Clone on the server (a read-only deploy key is enough):

   git clone <this repo> /opt/inneroracle && cd /opt/inneroracle

3. Copy the prebuilt data over (it is not in git):

   scp -r data/verses.db data/qdrant_local data/bm25.json your-server:/opt/inneroracle/data/

4. Configure — cp .env.example .env, then set:

   • ANTHROPIC_API_KEY
   • QDRANT_URL= — leave empty: the app runs Qdrant embedded from data/qdrant_local, so no Qdrant container is needed (the compose qdrant service sits behind the server-index profile and stays off).
   • TUNNEL_TOKEN — create a tunnel in Cloudflare Zero Trust → Networks → Tunnels, point its public hostname (your apex domain) at http://app:8000, and paste its token.

5. Start it:

   docker compose up -d --build app cloudflared

   The app binds to localhost only; the Cloudflare tunnel is the sole way in. Smoke-test: your domain loads, a question streams an answer, citations open the shloka, and https://your-domain/health returns {"status":"ok"}.

6. Protect the edge — the app itself ships with no login, so put the guardrails at Cloudflare: enable Bot Fight Mode, add a rate-limiting rule on /chat, and set a monthly spend cap in the Anthropic console.

7. Updating later:

   cd /opt/inneroracle && git pull && docker compose up -d --build app cloudflared

   Re-run step 1 + 3 only when the texts or the index change. Logs: docker compose logs -f app.

---

_{Made with care for the texts. Answers are for reflection, not a substitute for a teacher.}