A search system for everything on your disk.

// ingest

One pipeline. Six modalities.

A file's modality picks its preprocessor. A profiler picks its chunker. Each chunk is enriched in parallel, embedded into dense and sparse vectors from a single model, and upserted into a per-bucket Qdrant collection. If the subject is ambiguous (an unnamed photo, a document about an unnamed person), the model files clarifications for you to answer later.

1. aprofilemodality + structure. code, paper, image, audio, video, prose.
2. bpreprocessvision caption + OCR, whisper transcript, AST parse, or native text.
3. cchunkcode AST, paper-section, or prose. picked from the profile.
4. denrichsituating context + entity metadata, in parallel. images attached so visual context grounds in pixels.
5. eembedbge-m3 dense + sparse from one model. GPU when available.
6. findexqdrant named-vector upsert. postgres metadata. ambiguity flagged.

// answer

Retrieve, rerank, ground in the actual pixels.

A query rewrite tightens the search string. Qdrant runs dense and sparse prefetches and fuses them server-side with RRF, up to 20 candidates. The bge-reranker-v2-m3 cross-encoder scores the whole pool. A diversity guarantee makes sure every retrieved media modality lands its best chunk in the top-k. The prompt assembles citations; when one is an image or video, the original media is attached so vision-capable LLMs cite what is actually there. The answer streams as SSE deltas with 600 ms Postgres checkpoints, so a refresh, restart, or mid-stream crash never loses the buffer.

1. arewritequery transform. tightens search without changing intent.
2. bretrievedense + sparse prefetch. qdrant server-side RRF fusion. 20 candidates.
3. crerankbge-reranker-v2-m3 cross-encoder over the whole pool.
4. ddiversifyevery retrieved media modality earns one slot in the top-k.
5. ebuildcited spans + original media attached for vision-capable LLMs.
6. fstreamSSE deltas. 600 ms postgres checkpoints. real Stop kills the subprocess.

// why local

Some things should not be a SaaS.

Research notes. Source code. Private photos. The RAG that reads them should not be a billed monthly subscription with a privacy policy you skim. AdaRag owns its index, runs on one workstation, and routes inference through whichever local CLI you already pay for.

// stack

One job each.

• embeddings
  bge-m3dense + sparse from one model. 1024d. local on GPU when available.
• vector store
  Qdrantnamed dense + sparse vectors. server-side RRF prefetch fusion.
• rerank
  bge-reranker-v2-m3cross-encoder over the full candidate pool, every query.
• metadata
  Postgresdocuments, chunks, clarifications, chat jobs. async via asyncpg.
• api
  FastAPIPython 3.12. SQLAlchemy 2 async. background tasks for non-blocking ingest.
• frontend
  Next.js 16App Router + Turbopack. React 19. Tailwind v4. SSE via ReadableStream.
• llm bridge
  host CLIuvicorn on the host. proxies claude-cli, gemini-cli, ollama over HTTP.
• optimizer
  Optunamulti-objective. Pareto front of relevance vs latency over golden queries.