Limits and remaining deployment work

The first beta guide documents the current evidence/publication, operation, scoped-access and runtime contracts. It supplements the compatibility interfaces described below.

The tested scope is a local-first, single-corpus Rust application/library with optional HTTP model services and an authenticated API. Separate configs/processes support different corpora/models. Named principals have path-scoped access. It is not a multi-tenant identity service: OAuth/OIDC, source-defined ACLs, shared-corpus routing and automatic cross-model alignment are absent.

  • Search and compression benchmarks use small fictional fixtures. They validate integration/regressions, not large-corpus scalability or general model quality. sqlite-vec/experimental TurboQuant scanning is not an ANN deployment. Full vectors/cache remain stored, so compressed packets do not reduce total DB size in this mode.
  • Earlier validation recorded OCR and generation failures. Encryption/authentication and retrieval improvements do not eliminate hallucination. Generated entities/relations require source grounding; valid citation IDs do not prove an answer follows from its sources. Enrichment is disabled in the default native profile.
  • Video/GIF ingestion samples frames and segments audio. It does not recognize continuous motion/actions or ingest livestreams/HLS/YouTube player URLs. Audio diarization and TTS are absent. Raw multimodal embeddings are not transcripts or image descriptions.
  • PDF reading order, tables, bad font encodings and scan OCR depend on the decoder/models. Scanned pages need rendering plus vision/OCR for textual RAG. HTML capture is static UTF-8 extraction, not a browser mirror; JS, CSS/fonts/scripts, authenticated navigation, redirects and srcset are not mirrored.
  • Source/scan, decoded output and HTTP budgets are bounded, while query snapshots and individual decoder inputs/uploads are held in memory. Folder scans stage one file at a time. Native decoders/models are trusted host code, not security sandboxes. Set host memory/CPU/storage limits for untrusted large corpora.
  • SQLCipher covers database/WAL pages. Editable OKF files, temporary decoder input, key files, model memory, old plaintext reports/databases, swap and backups need their own protection. Unix file-mode checks are not Windows ACL enforcement. Encrypted backup/restore and explicit retention/GC are provided. Key rotation, OS keychain integration and scheduled retention policy are not automated.
  • Delete removes current source/active search data after reconciliation. History, caches and backups remain. One-file writes and capture-folder publication are atomic, but multi-file edits across independent external applications are not transactional. SHA-256 preconditions and a per-Engine lock protect cooperative API/library writers; external editors/processes can still race a final check/rename.
  • Capture deletion is restricted to app-managed captures with a current tree hash and supported regular files. A marker/tree policy failure requires inspecting/correcting the folder or removing files individually. Navigation links are preserved; deletion can leave dangling references.
  • Model HTTP protocols are explicit contracts, not universal provider adapters. Generic remote media encoders require a declared compatible shared space. Arbitrary provider request/response shapes need a Models adapter. Different vector spaces are not automatically ensembled together.
  • Linux/Windows/other accelerator execution requires the supplied CI/hardware checks. A build matrix is not evidence of a successful run until it executes. Current reports distinguish actual M3 execution, Intel compile checks and unexecuted target configurations.

Priorities for larger deployments include streaming/isolated large-media ingestion, automated retention and operational restore qualification, a scalable vector index, provider-specific adapters/evaluations, finer access controls, observability and broader hardware CI. These are documented future work rather than implied current capabilities. The requested implementation checklist is here.

Read Markdown →