v0.3.0 — CJK search, RRF fusion, LLM search hooks, ontology-backed agents
·
33 commits
to main
since this release
v0.3.0 — CJK search, RRF fusion, LLM search hooks, ontology-backed agents
Ontology search engine (Track A)
- Korean/CJK search now works. CJK bigram tokenizer + FTS5
trigramindex; existing databases migrate and re-embed automatically on first open (schema v2). - RRF hybrid ranking replaces mixed-scale fixed-weight fusion, computed over a bounded candidate pool (no full-corpus Python cosine scan).
- Host-LLM search hooks (optional, zero extra cost): query expansion and rerank callables injectable by the host CLI runtime (Claude Code / Codex). No embedding API or key required.
- Data-sovereignty gate: chunks scoped
private/confidentialare never passed to cloud rerank hooks — enforced inside the search pipeline and covered by a behavioral test. - 15% chunk overlap so context is not cut at window boundaries.
Builder wiring (Track B)
- New
ontology-backed-agentoverlay mode: corpus-dependent requests generate retrieval-first, source-cited agents (modes/ontology-backed-agent.md). - Rule-based governance contract injection via
.agentlas/contract-injection-map.json— only the contracts matching the agent's task traits are injected (no more blanket 26-contract packages). - Golden-path reference:
examples/ontology-proposal-agent/with a behavioralverify.sh(ingest → retrieve → draft → separate-context verify → human gate).
Architecture (Track C)
scripts/sync-adapters.shrenders the canonical core into runtime adapter mirrors;--checkis wired intoverify-package.shand blocks drift.loop_policy(none / self-correct / verified) derived from task risk tier; verified work uses a separate-context verifier (no self-grading).scripts/verify-mcp-surface.shguards theagentlasMCP registration contract across Claude Code, Codex, Gemini, and Antigravity.bump-version.shno longer hard-codes a local web path (usesAGENTLAS_WEB_INSTALL_GUIDE).
READMEs updated in en/ko/ja/zh-CN/hi. All verification gates pass: 24 unit tests, verify-package.sh, public_safety_check.sh, MCP surface check, adapter drift check, example agent behavioral verify.