jairo/Luma
Watch
1
0
Fork
You've already forked Luma
 0
Code Issues Pull requests Projects Releases 35 Packages Wiki Activity Actions
35 releases 38 tags
  • v4.8.0 ae96353374

    v4.8.0 Stable

    github-actions[bot] released this 2026-07-15 07:31:52 +00:00 | 30 commits to main since this release

    Full Changelog: https://github.com/Jairodaniel-17/Luma/compare/v4.7.0...v4.8.0

    Downloads
  • v4.7.0 11b92b20de

    v4.7.0 Stable

    github-actions[bot] released this 2026-07-15 01:14:24 +00:00 | 32 commits to main since this release

    Full Changelog: https://github.com/Jairodaniel-17/Luma/compare/v4.6.0...v4.7.0

    Downloads
  • v4.6.0 f66bd6d2ff

    v4.6.0 Stable

    github-actions[bot] released this 2026-07-15 00:33:19 +00:00 | 34 commits to main since this release

    Full Changelog: https://github.com/Jairodaniel-17/Luma/compare/v4.5.0...v4.6.0

    Downloads
  • v4.5.0 f0c16a7e62

    v4.5.0 Stable

    github-actions[bot] released this 2026-07-14 23:02:54 +00:00 | 36 commits to main since this release

    Full Changelog: https://github.com/Jairodaniel-17/Luma/compare/v4.4.0...v4.5.0

    Downloads
  • v4.4.0 442f4ee554

    v4.4.0 Stable

    github-actions[bot] released this 2026-07-14 20:30:46 +00:00 | 44 commits to main since this release

    Full Changelog: https://github.com/Jairodaniel-17/Luma/compare/v4.3.0...v4.4.0

    Downloads
  • v4.3.0 eae87ba78c

    v4.3.0 Stable

    github-actions[bot] released this 2026-07-14 14:03:47 +00:00 | 50 commits to main since this release

    Full Changelog: https://github.com/Jairodaniel-17/Luma/compare/v4.2.0...v4.3.0

    Downloads
  • v4.2.0 7ee6c04d53

    v4.2.0 Stable

    github-actions[bot] released this 2026-06-20 05:52:27 +00:00 | 86 commits to main since this release

    Full Changelog: https://github.com/Jairodaniel-17/Luma/compare/v4.1.1...v4.2.0

    Downloads
  • v4.1.0 36e6638385

    v4.1.0 — Object storage (S3-like blobs) Stable

    Jairodaniel-17 released this 2026-06-20 04:58:49 +00:00 | 89 commits to main since this release

    ✨ Nuevo: módulo de almacenamiento de objetos tipo S3 (blobs)

    Luma ahora puede guardar objetos binarios en disco, fuera del document store — ideal para adjuntos sin meter base64 en JSON.

    Endpoints (autenticados con Authorization: Bearer <api_key>)

    • PUT /v1/blob/:bucket/:key — sube bytes crudos (body binario). Devuelve { bucket, key, size, etag }.
    • GET /v1/blob/:bucket/:key — descarga los bytes.
    • DELETE /v1/blob/:bucket/:key — borra (idempotente).
    • GET /v1/blob/:bucket — lista las keys del bucket.

    Los objetos se persisten en {data_dir}/blobs/{bucket}/{key} con escritura atómica (tmp + rename), validación anti path-traversal (charset + componentes + prefix-check), límite por max_body_bytes y etag (crc32).

    Binario precompilado

    • luma-v4.1.0-linux-x86_64 — Linux x86-64, enlazado dinámicamente a glibc (≥ 2.x). Ejecuta: 
      chmod +x luma-v4.1.0-linux-x86_64
LUMA_API_KEY=tu-clave PORT=1234 ./luma-v4.1.0-linux-x86_64 serve
    • Verifica la integridad con el .sha256 adjunto: sha256sum -c luma-v4.1.0-linux-x86_64.sha256.

    Incluye también la config recomendada para escritura rápida (wal_sync_mode = "group").

    Downloads
  • v3.0.0 a512346de3

    v3.0.0 Stable

    Jairodaniel-17 released this 2026-04-22 21:07:17 +00:00 | 99 commits to main since this release

    CHANGELOG

    🚀 Novedades en v3.0.0 (Search, Observabilidad y NS-Mem Avanzado)

    Búsqueda y exportación

    • Batch search (POST /v1/vector/{collection}/search_batch): hasta 100 queries ejecutadas en paralelo internamente (rayon). Reduce 100 round-trips a 1.
    • Scroll / cursor API (GET /v1/vector/{collection}/scroll): paginación lexicográfica con cursor opaco para exportar colecciones completas sin límite de k.
    • Reranking por coseno (POST /v1/vector/{collection}/rerank): recibe IDs + query (texto o vector), embebe si es texto, reordena por coseno real. Ideal para pipeline search-then-rerank.
    • Aggregations (POST /v1/vector/{collection}/aggregate): { "group_by": "campo", "filter": {...}, "limit": N } — cuenta ítems por valor usando el keyword index. Fast path O(1) para campos indexados.
    • Pre-filter threshold configurable (pre_filter_threshold, default 10 000): brute-force automático sobre subconjuntos filtrados, más eficiente que HNSW + post-filter para corpus muy filtrados.

    Embeddings

    • 4 nuevos providers: azure (Azure OpenAI), cohere (v1/embed con input_type), huggingface (Inference API), añadidos a los existentes openai, ollama, mock.
    • Retry con backoff exponencial + jitter: EMBEDDING_RETRY_ATTEMPTS (default 3) y EMBEDDING_RETRY_INITIAL_MS (default 200). Resiste 429 y 503 transitorios sin fallar la request.

    Observabilidad y operaciones

    • Audit log (GET /v1/admin/audit): cada request queda registrada en SQLite con ts, api_key_id, ip, method, path, status, latency_ms. Filtra por rango de tiempo, key e id.
    • Backup endpoint (POST /v1/admin/backup): dispara snapshot WAL y retorna { "ok": true, "offset": N }. Ambos requieren rol admin.
    • Bench CI: cargo bench --no-run como job en GitHub Actions — bloquea merges que rompen compilación de benchmarks.
    • RBAC tests (tests/auth_rbac.rs): 7 tests de integración que cubren 401/403 en todas las combinaciones de token ausente, inválido, revocado y rol user vs admin.

    NS-Mem — Memoria de agentes más inteligente

    • Deduplicación de facts: el consolidador verifica similitud semántica (cosine ≥ 0.95) antes de insertar. Facts redundantes extraídos de eventos distintos no se acumulan.
    • Decay exponencial (memory_decay_enabled): campo decay_score en cada fact semántico. Decae con semivida configurable (memory_decay_half_life_days, default 30d). Facts por debajo del umbral se archivan automáticamente.
    • Detección de contradicciones: al sobrescribir un fact (upsert_fact), si la similitud semántica entre el contenido viejo y el nuevo es < 0.55, se crea una arista Contradicts en lugar de Supersedes y se emite log de contradicción detectada.

    v2.1.0 (2026-04-22)

    Added

    • Typed MetadataFilter system (src/vector/filter.rs): composable filter tree replacing the flat {"field": "value"} object.
      • Operators: eq, neq, gt, gte, lt, lte, in, not_in, any_of, contains, starts_with, exists.
      • Logical combinators: and, or, not (arbitrarily nested).
      • New filter field on SearchOptions (typed); legacy filters field kept for backward compatibility — both are merged with AND when both are present.
    • any_of operator for array-valued metadata fields: returns true when the field (JSON array) contains at least one of the query values. Example: tax_system: ["suitetax","legacy"] matches any_of: ["suitetax"]. Supports multi-system queries (any_of: ["suitetax","v3"]).
    • Keyword index extended to array fields: add_meta_to_index now indexes each string element of array-valued fields individually, enabling the keyword fast-path for any_of and eq queries on array fields without a full scan.
    • Keyword index fast path for any_of: resolves candidate ID sets directly from the index (union of all query values), same O(1) lookup as eq.
    • SQL translation for hub pre-filtering (to_sql_where): any_of maps to EXISTS (SELECT 1 FROM json_each(metadata, '$.field') WHERE value IN (...)).
    • from_legacy(): automatic conversion of old flat-object filters to typed And([Eq(...)]) — no breaking change for existing API clients.

    Internal

    • Removed matches_filters() free function and Collection::keyword_candidates() method; replaced by filter::evaluate_filter() and filter::index_candidates().
    • 7 new tests for any_of covering: single-value match, multi-value OR semantics, scalar-field non-match, missing field, keyword index fast path, AND combination, SQL translation.

    v2.0.1 (2026-04-22)

    Security

    • CVE fix: bump rustls-webpki 0.103.12 → 0.103.13 (RUSTSEC-2026-0104 — reachable panic in CRL parsing via DoS).
    • Remove API key support via query parameter — keys were leaking into access logs and proxy logs.
    • Enforce admin role check on all /v1/auth/keys endpoints; previously any valid key could create or revoke keys.

    Robustness

    • Switch EmbeddingClient and Metrics from std::sync::Mutex to parking_lot::Mutex — eliminates mutex-poison cascade panics.
    • Add bounded concurrency semaphore to EmbeddingClient (with_limits) — caps inflight HTTP requests to embedding providers.
    • Chunk OpenAI embed_batch into ≤ 96 texts per request to avoid rate limits and token overflow.
    • Log TriggeredBy edge failures in consolidator with tracing::warn! instead of silently swallowing errors.
    • Raise default HNSW compaction tombstone ratio from 0.2 → 0.5 (conservative by default, avoids compaction under normal write load).

    CI

    • Remove continue-on-error: true from test and fmt jobs — failing tests now block merges.
    • Add MSRV check job (Rust 1.75).
    • Add cargo-deny job with deny.toml for license, source, and advisory enforcement.

    v2.0.0 (2026-04-20)

    Added

    • NS-Mem Graph Layer: replaced flat K-NN recall with semantic walk BFS over typed memory graph (memory_edges).
      • Recall score: cosine × edge_factor × (1 + PageRank).
      • Edge factors: supports=1.0, triggered_by=0.8, related_to=0.7, contradicts=-0.5 (skip), supersedes=0.0 (skip).
    • memory_edges table: typed weighted edges with auto-creation on consolidation (TriggeredBy) and on upsert_fact overwrite (Supersedes).
    • memory_history table: append-only belief versioning — every upsert_fact on an existing fact snapshots the old version.
    • Simplified PageRank (15 iterations, damping 0.85) stored in memory_records.centrality_score; recomputable via POST /v1/memory/{namespace}/graph/centrality.
    • 5 new API endpoints:
      • POST /v1/memory/{namespace}/edges — create/update edge
      • GET /v1/memory/{namespace}/edges/{memory_id} — list edges for a node
      • POST /v1/memory/{namespace}/edges/{edge_id}/delete — remove edge
      • GET /v1/memory/{namespace}/beliefs/{fact_key}/history — belief version history
      • POST /v1/memory/{namespace}/graph/centrality — trigger PageRank recomputation

    Security

    • Updated quinn-proto → 0.11.14, rustls-webpki → 0.103.12.

    CI

    • All checks pass: fmt, clippy -D warnings, 99/99 tests, 0 audit errors.

    v1.4.0 (2026-03-15)

    Added

    • NS-Mem (src/memory/): episodic / semantic / procedural / working memory layer.
      • Consolidation pipeline: ingest_event → LLM fact extraction → semantic promotion.
      • Procedural memory: DAG-based procedures with typed edges, priority ordering, and constraint evaluation.
      • LLM providers: none, mock, openai, ollama.
      • Full REST API at /v1/memory/{namespace}/.
    • WAL hardening: checksummed envelopes, corruption stop-at-tail semantics, idempotent recovery for state_db and vector replay.
    • Range scans for KV: start/end parameters (end-exclusive) on GET /v1/state.
    • Cached metrics: RSS read outside the scrape path; GET /v1/metrics returns pre-rendered Prometheus text.
    • SSE gap signaling: explicit gap events when clients request offsets no longer retained.

    Documentation

    • docs/NS_MEM.md — full API reference for the memory layer.

    v1.3.0 (2025-12-20)

    Added

    • LumaDatabase hub (/v1/db/{namespace}/): orchestrates auto-chunking, embedding generation, hybrid SQL+vector search, and auto-schema indexing.
    • Hybrid search planner: decides sql_first vs vector_first strategy based on filter selectivity and collection size.
    • Multi-tenant isolation: per-namespace vector collections and SQL tables.
    • DiskANN index: disk-based Vamana graph for collections beyond HNSW memory limits. CLI: diskann-build, diskann-status.

    v1.2.0 (2025-10-05)

    Added

    • IVF_FLAT_Q8 index: inverted file index with 8-bit quantization refinement for large-scale approximate search.
    • SIMD acceleration: AVX2 dot-product and int8 dot-product intrinsics with runtime feature detection.
    • mmap vector storage: zero-copy memory-mapped access to frozen segments.
    • Segment compaction: background compaction triggered by tombstone ratio threshold.

    v1.1.0 (2025-08-15)

    Added

    • HNSW segmentation: collections split into ~8,192-vector segments; active segment receives upserts, frozen segments are read-only. Search merges results across all segments.
    • Pub/Sub events (/v1/events, /v1/stream): SSE-based event bus with monotonic offsets and gap detection.
    • TTL for KV state: per-key TTL with background expiration.
    • Optimistic locking: if_revision compare-and-swap for KV put/delete.
    • OpenAPI spec (docs/openapi.yaml): generated from route definitions.

    v0.2.0 (2025-12-14)

    Added

    • Secure bind (127.0.0.1 default; --bind/--unsafe-bind flags for explicit exposure).
    • Consistent error responses ({ "error": "...", "message": "..." } across all routes).
    • Batch APIs: /v1/state/batch_put, /v1/vector/*_batch.
    • DocStore over KV (/v1/doc/* + metadata find).
    • SQLite module (/v1/sql/query, /v1/sql/exec).
    • Vector store segmentation + metadata keyword index.
    • vacuum CLI command for vector collection compaction.

    v0.1.0 (2025-10-01)

    Initial release. Core primitives:

    • Vector store with HNSW (in-memory, single segment).
    • Key-value state store with JSON values.
    • WAL persistence with periodic snapshots.
    • Axum HTTP server with Bearer token auth.
    • Basic cargo bench for vector add/search.
    Downloads
  • v2.0.0 0ffa7e016d

    v2.0.0 Stable

    Jairodaniel-17 released this 2026-04-16 20:50:11 +00:00 | 105 commits to main since this release

    Downloads