root 5a3364f539 matrix: judge-gated Shape B inject — closes lift-suite tail issues
Lift suite run #004 left two unresolved tail issues:
- Q6 ("Forklift loader") ↔ Q7 ("Hazmat warehouse, cold storage")
  swap recordings as warm top-1 because their embeddings are within
  0.20 cosine of each other. Distance gate can't tell them apart.
- Q9 + Q15 lose paraphrase recovery when qwen2.5 rephrases past the
  0.20 threshold. Distance says "drift too far"; sometimes the drift
  is real (skip), sometimes the paraphrase is still on-domain (don't
  want to skip).

Multi-coord run #008's judge re-rating proved the LLM can
distinguish: Q3 crane case landed at distance 0.23 (looks tight)
but rating 1 (irrelevant). The judge sees domain mismatch the
embedder doesn't.

This commit lifts that pattern into the matrix substrate. Shape B
inject now optionally routes every candidate through a judge gate
before the rank insert lands. Distance + judge BOTH have to approve.

internal/matrix/playbook.go:
- InjectPlaybookMisses signature gains a query string + an
  optional InjectGate. nil gate preserves pre-judge-gating
  behavior (current tests already pass with nil).
- New InjectGate interface + InjectGateFunc adapter for tests
  and non-LLM callers.
- Per-candidate gate.Approve(query, hit) call inserted between
  the dedup and the inject. Rejected candidates skip silently;
  injected count reflects post-gate decision.

internal/matrix/judge.go (new, ~140 lines):
- LLMJudgeGate calls an Ollama-shape /api/chat endpoint with the
  same 1-5 staffing-rubric prompt that worked in multi_coord
  run #008. fail-closed on HTTP/JSON errors (don't inject if
  judge can't speak — better miss than wrong-domain).
- NewLLMJudgeGate returns nil when URL or Model is empty,
  matching InjectGate's nil-means-no-judge semantics.

internal/matrix/retrieve.go:
- SearchRequest gains JudgeURL, JudgeModel, JudgeMinRating
  fields. Run() builds an LLMJudgeGate when set; passes nil
  otherwise. Backward compatible — existing callers see no
  behavior change.

Tests:
- TestInjectPlaybookMisses_GateRejectsCandidate (rejectAll → 0
  injected, even with tight distance)
- TestInjectPlaybookMisses_GateApprovesCandidate (approveAll →
  same as nil-gate behavior)
- TestInjectPlaybookMisses_GateSeesCorrectQuery (gate receives
  CURRENT query + RECORDED query separately so it can score
  the (current, candidate) pair)
- All 5 existing inject tests updated to new signature

go test ./internal/matrix → all 8 inject tests pass.
go test ./internal/matrix ./internal/shared ./cmd/{matrixd,
queryd,pathwayd,observerd} → all green.

STATE_OF_PLAY:
- OPEN item #1 (judge-gated injection) closed.
- DO NOT RELITIGATE adds the substrate-level judge-gate lock.
- OPEN list now 5 rows (was 6).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-30 19:38:12 -05:00

golangLAKEHOUSE

Go reimplementation of the Lakehouse — a versioned knowledge substrate for staffing analytics + local AI workloads.

Status

Phase G0 complete + G1/G1P/G2 shipped. Six binaries plus a seventh (vectord) and an eighth (embedd) on top, fronted by a single gateway. Acceptance smokes green for D1-D6 + G1 + G1P + G2.

End-to-end staffing co-pilot pipeline functional through the gateway:

text → /v1/embed → /v1/vectors/index/<name>/add
text → /v1/embed → /v1/vectors/index/<name>/search → top-K hits

Plus the SQL path:

CSV  → /v1/ingest    (parses, writes Parquet via storaged, registers
                      manifest with catalogd)
SQL  → /v1/sql       (DuckDB over the registered Parquets via httpfs)

See docs/PHASE_G0_KICKOFF.md for the day-by-day record (D1-D6 + real-scale validation + G1/G1P/G2 pointer at the bottom).

Service inventory

Bin Port Role
gateway 3110 Reverse proxy fronting all backing services
storaged 3211 Object I/O over S3 (MinIO in dev)
catalogd 3212 Parquet manifest registry, ADR-020 idempotency
ingestd 3213 CSV → Parquet → register loop
queryd 3214 DuckDB SELECT over registered Parquets via httpfs
vectord 3215 HNSW vector search (+ optional persistence to storaged)
embedd 3216 Text → vector via Ollama (default nomic-embed-text 768-d)
mcpd stdio Model Context Protocol server (Claude Desktop / Code consumers)

MCP server

bin/mcpd exposes Lakehouse capabilities as MCP tools over stdio: list_datasets, get_manifest, query_sql, embed_text, search_vectors. All tools proxy to the gateway, so the gateway must be up first.

Wire into Claude Desktop / Claude Code by adding to the MCP config:

{
  "mcpServers": {
    "lakehouse": {
      "command": "/path/to/golangLAKEHOUSE/bin/mcpd",
      "args": ["--gateway", "http://127.0.0.1:3110"]
    }
  }
}

Replaces the Bun mcp-server.ts MCP-tool surface from the Rust system. HTTP demo routes (the staffing co-pilot UI) stay Bun until G5.

Acceptance smokes

scripts/d1_smoke.sh   # 5-binary skeleton + chi /health + gateway proxy probes
scripts/d2_smoke.sh   # storaged GET/PUT/LIST/DELETE + 256 MiB cap + concurrency cap
scripts/d3_smoke.sh   # catalogd register/manifest/list + rehydrate-across-restart
scripts/d4_smoke.sh   # ingestd CSV → Parquet round-trip + schema-drift 409
scripts/d5_smoke.sh   # queryd DuckDB SELECT through httpfs over MinIO
scripts/d6_smoke.sh   # full ingest → query through gateway only
scripts/g1_smoke.sh   # vectord HNSW recall + dim mismatch + duplicate-create 409
scripts/g1p_smoke.sh  # vectord state survives kill+restart via storaged
scripts/g2_smoke.sh   # embed → vectord add → search round-trip

Or run the full gate via the task runner (see below):

just verify     # vet + tests + 9 smokes; ~33s wall

Task runner

just                 # show available recipes
just verify          # full Sprint 0 gate (vet + tests + 9 smokes)
just smoke <day>     # single smoke (d1..d6, g1, g1p, g2)
just doctor          # check cold-start deps; --json for CI
just install-hooks   # install pre-push hook that runs just verify

After a fresh clone, run just install-hooks once so git push is gated on the same green chain that ran here. Hook lives in .git/hooks/pre-push (not tracked; recreated by the recipe).

Cold-start dependencies

  • Go 1.25+ at /usr/local/go/bin (arrow-go pulled the 1.25 floor)
  • gcc + libc-dev for the DuckDB cgo binding (ADR-001 §1.1)
  • just task runner (apt install just on Debian 13+)
  • MinIO running on :9000 with bucket lakehouse-go-primary
  • Ollama running on :11434 with nomic-embed-text loaded (G2)
  • /etc/lakehouse/secrets-go.toml with [s3.primary] credentials (storaged + queryd both read this)

just doctor probes all of the above and reports the fix command for each missing dep. CI / scripts can use just doctor --json.

Layout

docs/                         Direction + spec + ADRs + day-by-day
cmd/                          One main package per binary
internal/                     Shared packages — storeclient, catalogclient,
                                secrets, shared, embed, gateway, plus
                                per-service implementation packages
scripts/                      Smokes + ancillary tooling

Reading order

  1. docs/PRD.md — what we're building and why
  2. docs/SPEC.md — how, per-component
  3. docs/DECISIONS.md — ADRs (ADR-001 foundational)
  4. docs/PHASE_G0_KICKOFF.md — day-by-day from D1 through G2
  5. docs/RUST_PATHWAY_MEMORY_NOTE.md — historical reference for the Rust era's pathway memory (not migrated, by ADR-001 #5)

Predecessor

The Rust Lakehouse this rewrite supersedes lives at git.agentview.dev/profit/lakehouse. It remains the live system serving devop.live/lakehouse/ until this Go implementation reaches feature parity per docs/SPEC.md §7. Then Rust enters maintenance-only mode.

Description
Go reimplementation of the Lakehouse — versioned knowledge substrate for staffing analytics + local AI workloads
Readme 3.2 MiB
Languages
Go 79.4%
Shell 20.1%
Just 0.3%
Dockerfile 0.2%