profit/lakehouse
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
39 Commits 13 Branches 3 Tags
Commit Graph

5 Commits

Author SHA1 Message Date
root
04770c97eb HNSW vector index: 100K search in 27ms (58x faster than brute-force) 
- instant-distance HNSW implementation for approximate nearest neighbors
- HnswStore: build from stored embeddings, in-memory index, thread-safe
- POST /vectors/hnsw/build — build index from Parquet (100K in 35s release)
- POST /vectors/hnsw/search — fast ANN search
- GET /vectors/hnsw/list — list loaded indexes

Benchmark (100K × 768d, release build):
  Brute-force: 1,567ms
  HNSW:           31ms (50x)
  HNSW warm:      27ms (58x)

Build cost: 35s one-time for 100K vectors (release mode)
ef_construction=40, ef_search=50 — good recall/speed balance

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-03-27 20:00:50 -05:00
root
6cd1daeb51 Phase 11: Embedding versioning — model-proof vector layer 
- IndexRegistry: tracks all vector indexes with model metadata
  (model_name, model_version, dimensions, build stats)
- Index metadata persisted as JSON in vectors/meta/
- Rebuilt on startup for crash recovery
- GET /vectors/indexes — list all indexes (filter by source/model)
- GET /vectors/indexes/{name} — get index metadata
- Background jobs auto-register metadata on completion
- Multi-version support: same data, different models, coexist
- Per ADR-014: enables incremental re-embed on model upgrade

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-03-27 09:27:10 -05:00
root
3b695cd592 Dual-pipeline supervisor for embedding ingestion 
- 4 parallel pipelines (tuned for i9 + A4000)
- Range-based work splitting (2500 chunks per range)
- Round-robin retry on failure (3 attempts before dead-letter)
- Checkpointing to disk every 1000 chunks (crash recovery)
- On restart, loads checkpoint and skips completed ranges
- Dead-letter queue for permanently failed ranges
- Vectors assembled in order after all pipelines finish
- Batch size 64 for GPU throughput

Architecture:
  Supervisor → splits 100K chunks into 40 ranges
    ├── Pipeline 0: grabs range, embeds, reports progress
    ├── Pipeline 1: grabs range, embeds, reports progress
    ├── Pipeline 2: grabs range, embeds, reports progress
    └── Pipeline 3: grabs range, embeds, reports progress
    Failed range → back to queue → next available pipeline retries

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-03-27 09:06:28 -05:00
root
6a532cb248 Background job system for embedding — fixes 100K timeout 
- JobTracker: create/update/complete/fail jobs with progress tracking
- POST /vectors/index now returns immediately with job_id (HTTP 202)
- Embedding runs in tokio::spawn background task
- GET /vectors/jobs/{id} returns live progress (chunks embedded, rate, ETA)
- GET /vectors/jobs lists all jobs
- Progress logged every 100 batches with chunks/sec and ETA
- 100K embedding job running successfully at 44 chunks/sec
- System stays responsive during embedding (queries in 23ms)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-03-27 09:03:07 -05:00
root
26fc98c885 Phase 7: Vector index + RAG pipeline 
- vectord crate: chunk → embed → store → search → RAG
- chunker: configurable chunk size + overlap, sentence-boundary aware splitting
- store: embeddings as Parquet (binary blob f32 vectors), portable format
- search: brute-force cosine similarity (works up to ~100K vectors)
- rag: full pipeline — embed question → search index → retrieve context → LLM answer
- Endpoints: POST /vectors/index, /vectors/search, /vectors/rag
- Gateway wired with vectord service
- Tested: 200 candidate resumes indexed in 5.4s, semantic search + RAG working
- 20 unit tests passing (chunker, search, ingestd, shared)
- AI gives honest "no match found" when context doesn't support an answer

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-03-27 08:12:28 -05:00