revvec

Airgapped industrial RAG. On device. No cloud.

🌐 revvec.vercel.app  ·  📺 3 minute demo

Submission for the Actian VectorAI DB Build Challenge, April 2026.

Why this exists

Defense, aerospace, pharma, semiconductor, EU critical infrastructure. None of them can ship their SOPs, sensor data, or mission docs to a cloud LLM. CMMC, ITAR, 21 CFR Part 11, NIS2 say no.

revvec runs every model and every byte of your data on a single Mac. Pull the ethernet cable and it still answers, still cites, still works.

Architecture

Five stages, one Mac, zero hops. Inputs (PDFs, images, sensor windows, voice) go through five embedding models, all named vectors land in one Actian collection, retrieval is a three tier hybrid (server side RRF prefetch + Okapi BM25 rerank + answer cache), and the response streams from Qwen3-4B with [source:N] citations.

What ships

• Mac .app with multi chat sidebar, streaming token answers, live voice transcription, inline [source:N] citation pills that open the cited PDF at the cited page.
• FastAPI sidecar running locally on 127.0.0.1:8000. Every endpoint is async def because MLX GPU streams are thread local on macOS.
• Actian VectorAI DB in Docker on localhost:50052. One collection revvec_memory. Three named vectors per point.
• Hash chained audit log (SHA-256 over canonical JSON, JSONL per UTC day). 6 of 6 tamper tests pass.
• NASA Mars 2020 corpus. Public domain, ITAR plausible, abundant.

Screenshots

A grounded answer with the four engineer personas, the source PDF panel sliding in on the right, and the Actian inspector dropdown at the bottom of every answer.

The Actian inspector strip. One click expands it and surfaces every Actian primitive the query exercised. Vector search path, vectors used, hits returned with top score, latency split between Actian and the BM25 rerank, the actual operation as pseudocode, and a chip cloud of the Actian features in play.

Citations resolve to real entity_ids. Click an orange pill and the source PDF opens at the cited page with the cited phrase highlighted.

Live fusion mode toggle. Every query is routed through points.query(prefetch=[...], query={fusion: ...}). Switch between RRF and DBSF mid-session and the Inspector reflects the mode you picked.

Run it

make install     # Python 3.12 via uv (one time)
make up          # start Actian in Docker
make serve       # start FastAPI sidecar (blocks)

# in another terminal
make app         # opens the bundled .app

Built artifacts after a release build:

• app/src-tauri/target/release/bundle/macos/revvec.app
• app/src-tauri/target/release/bundle/dmg/revvec_0.1.0_aarch64.dmg (about 3.1 MB)

Stack

Slot	Model
LLM	Qwen3-4B-Instruct-2507-4bit via MLX
Text embed	Qwen3-Embedding-0.6B (1024d)
Image embed	DINOv2-large (1024d)
Sensor embed	Chronos-Bolt-small (512d)
ASR	Whisper-large-v3-turbo via mlx-whisper
TTS	Kokoro-82M
Vector DB	Actian VectorAI DB (Docker, port 50052)
Frontend	Tauri 2 + React 19 + Tailwind
Backend	FastAPI + Uvicorn (Python 3.12, uv)

All Apache-2.0 or ungated. No cloud API at runtime. Weights live in ~/.cache/huggingface/.

Numbers

P95 retrieval	68 ms
Cache hit P95	9 ms
Fabricated citations	0
Audit tamper tests	6 / 6
.dmg size	3.1 MB

Why Actian is load bearing

1. Three named vectors on one point, one shared payload. A PDF page becomes one record holding both a text vector and an image vector. ColPali inspired.
2. Server side RRF fusion via client.points.query(prefetch=...). One round trip across all named vectors. About 15 ms on a 1,940 point corpus.
3. set_payload for in place state transitions. Pattern promotion (candidate to active) without re-embedding the vector.
4. points.search(score_threshold=0.95) for the answer cache. One server call combines a cosine threshold with a payload filter. Cache hits in 9 ms.
5. FilterBuilder DSL for compliance flavoured queries. Forget by entity_id, scope by equipment, scope by time range.
6. HNSW tuned at create time (m=32, ef_construct=256). Schema is locked, never mutated.
7. vde.open_collection() on every ensure_ready(). Actian's beta collections.exists() lies after a server restart, so the writer always re-mounts. Robust to Docker bounces.
8. Industrial code aware Okapi BM25 reranker stitched on top of the Actian stage one pool. Tokenizer keeps codes like SOP-ME-112 and sol 1214 intact (codes embeddings have never seen). 0.7·cosine + 0.3·BM25 with a graceful fallback.

Take Actian out and you would need four databases plus a client side merge layer.

Architecture (code)

Nine single file agents under src/revvec/:

ingestion/orchestrator.py    fans to sop, image, sensor, voice, log
embed/service.py             lazy-loads all 5 models with TTL unload
memory/actian_writer.py      sole writer into Actian
cluster/promotion.py         candidate -> active state machine
retrieval/hybrid.py          RRF prefetch + BM25 rerank + answer cache
llm/qwen_mlx.py              Qwen3-4B via MLX
voice/stt_tts.py             Whisper + Kokoro + live streaming
audit/chain.py               hash chained JSONL
persona/router.py            persona specific prompt overlays

The web landing page lives in web/ (Next.js, deployed to Vercel).

Compliance

Hash chained audit log over canonical JSON. Every query, cache hit, and forget request gets recorded. Verify with GET /api/admin/audit. Tamper any byte and the next verify call returns chain_ok: false.

Maps to CMMC 2.0 SP.3.1, ITAR 22 CFR 120, 21 CFR Part 11 §11.10(e), NIS2 Art. 21, GDPR Art. 17. Honest about gaps. No auth on admin endpoints yet, and vde.save_snapshot returns UNIMPLEMENTED on the current Actian beta (the audit row is written either way).

License

Apache-2.0. Models keep their own licences.