Skip to content

Latest commit

 

History

History
31 lines (27 loc) · 3.08 KB

File metadata and controls

31 lines (27 loc) · 3.08 KB

HybridMind implementation status

Updated: August 2026

This is a code-status inventory, not a quality or production-readiness claim.

Area Current status Evidence boundary
Remote 4096-d embeddings Implemented, fail-closed Responses, batches, SQLite blobs, rebuilds, and queries validate finite exact width; live provider health still needs bounded preflight.
Dense retrieval Implemented FAISS HNSW candidate generation; real-corpus quality and scale are unproven.
Sparse retrieval Implemented BM25/BM25S independent candidate path.
Typed graph retrieval Implemented Directed parallel edges, edge IDs, confidence, validity and explicit anchors affect traversal; graph usefulness remains benchmark-dependent.
RRF fusion Implemented Validated numeric inputs and explicit-weight semantics.
Cross-encoder reranker Optional, off by default Search exposes attempted/applied/failure state; a positive evaluator pool must prove execution.
Query routing Implemented heuristic Regex classification changes omitted weights; it is not an LLM or learned router.
Query decomposition Optional Bounded centralized-LLM path with guards; improvement is unproven.
Temporal handling Partially implemented Event/valid/assertion times, half-open filtering, latest/previous and conservative supersession exist; this is not general temporal reasoning.
Structured facts Optional Provider-extracted fields, stable IDs, provenance and conservative slot updates; extraction accuracy requires evaluation.
Salience/access Optional heuristic Recency/access/degree multiplier; not learned and disabled by default.
Consolidation Optional lossy summary Provenance-linked derived summaries; not Observer/Reflector and sources are retained.
Portable snapshots Implemented v2 Safe allowlisted files, SQLite backup, integrity/semantic verification and staged restore; unsigned manifests do not prove authenticity.
API security Local-safe baseline Loopback default, remote-bind key requirement, CORS/host allowlists and process-local rate limits; not multi-tenant auth.
LoCoMo/LongMemEval/MuSiQue retrieval evaluators Hardened but awaiting fresh live runs Exact evidence metrics, scoping, failed receipts and manifests; historical legacy results are not valid current evidence.
Ablation matrix Partial Vector/sparse conditions can be request-attested; graph/hybrid/stateful conditions remain plan-only without external manifests/server attestation.
GNN/fusion training scripts Scaffolded/untrained No shipped trained checkpoints or validated gains.
ColBERT/visual retrieval Experimental/optional Not part of the default validated path.
10M–100M effective context Research target Protocol and capacity arithmetic exist; target quality has not been demonstrated.
Transformer KV-cache replacement Not implemented Requires model-integrated attention/KV access, not semantic retrieval alone.

Current test results should always be reported from a fresh full-suite run. Do not reuse counts from this document.