mirror of
https://github.com/garrytan/gbrain.git
synced 2026-07-30 11:22:34 +00:00
f004a27429
* docs(designs): 2026-05 embedder shootout eval plan Adds docs/designs/2026_05_EVAL_PLAN.md — the approved plan + 6 Conductor session briefs for the OpenAI vs Voyage vs ZeroEntropy embedder comparison. Why: produce a publishable comparison report for v0.35.x release notes pinning "which embedder wins, and does zerank-2 carry the win for ZeroEntropy" against public LongMemEval + in-house BrainBench. Each session brief is self-contained — repo, branch, commits, verify, ship, deliverable, hand-off. Stewardable one section per Conductor session. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(pricing): add voyage-4-large + zembed-1 to EMBEDDING_PRICING v0.35.0.0 shipped ZeroEntropy zembed-1 + zerank-2 reranker support and expanded the Voyage allow-list to include voyage-4-large. The pricing table missed both, so `gbrain upgrade`'s post-upgrade reembed prompt silently fell back to "estimate unavailable" for users on these models. - voyage:voyage-4-large @ $0.18/MTok (same as voyage-3-large) - zeroentropyai:zembed-1 @ $0.05/MTok New test file pins both entries plus the openai/voyage-3-large baselines, case-insensitive provider matching, bare-model openai-default fallback, table integrity (lowercase providers, finite non-negative prices), and the estimateCostFromChars approximation. 11 cases, 46 expect() calls. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(exports): expose gbrain/ai/gateway with canary test Adds ./ai/gateway to the package.json exports map so external eval consumers (notably gbrain-evals, the sibling repo running the embedder shootout in docs/designs/2026_05_EVAL_PLAN.md) can call configureGateway directly to swap embedding providers per cell. Why: pre-v0.35.1.0, gbrain-evals adapters hardcoded gbrain/embedding, which means every retrieval adapter was OpenAI-only. The newly-exposed gateway lets adapters route through Voyage and ZeroEntropy without forking gbrain or duplicating the recipe wiring. - package.json: add "./ai/gateway" -> "./src/core/ai/gateway.ts" - scripts/check-exports-count.sh: bump expected count 17 -> 18 - test/public-exports.test.ts: add canary pinning configureGateway + embed, bump expected count assertion Pre-existing import-resolution failures in this test file (16 on master) are unrelated to this change — they're a longstanding Bun package self-import behavior. The count + EXPECTED_EXPORTS list-match assertions both pass cleanly. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(eval): add --resume-from <jsonl> to gbrain eval longmemeval Multi-cell embedder shootouts spend $50+/cell on the gpt-4o judge after gbrain emits hypotheses. A mid-run abort (rate-limit, cost-cap, OS interrupt, SIGKILL) previously meant re-paying the full cell. This flag makes those aborts cheap: re-invoke with --resume-from pointed at the partial JSONL and only the unanswered question_ids re-run. Behavior: - Read question_ids from the file; skip them on this run. - Rows with non-empty hypothesis count as done. - Rows with hypothesis="" AND an error field are NOT skipped (retry case for per-question failures recorded by the existing try/catch). - Corrupt trailing lines (SIGKILL'd writer mid-line) are silently skipped with a stderr warn. - When --resume-from path == --output path, the output emitter opens the file in append mode instead of truncating, so the existing rows survive. - Empty resume case (all questions already done) returns immediately without spinning up the brain or calling the client. New exported helper loadResumeSet() makes the parser unit-testable. 6 new test cases pinning: - File-not-found returns empty set - Well-formed JSONL load - Error-row retry semantics (empty hypothesis + error -> not in set) - Truncated final line recovery - End-to-end resume against the 5-question mini fixture - All-done early-return (stub client must NOT be invoked) All 18 cases in test/eval-longmemeval.test.ts green; bun run typecheck clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore: v0.35.1.0 Bumps VERSION + package.json + CHANGELOG entry for the embedder-shootout prereq release. Three additive changes from the prior 4 commits: - pricing: voyage-4-large + zembed-1 entries - exports: gbrain/ai/gateway is now public - eval: gbrain eval longmemeval --resume-from <jsonl> Each commit on this branch is independently bisect-friendly and CI-green; the CHANGELOG entry is the user-facing rollup. No migrations, no breaking changes — the gateway export expands the surface, the resume-from flag is additive, the pricing patch only changes "estimate unavailable" -> a real dollar figure for two specific models. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(eval): longmemeval adapter handles _s split + sanitizes session_id slugs Three tightly-coupled bugs blocked `gbrain eval longmemeval` against the public LongMemEval _s split from HuggingFace (the dataset every shootout cell needs): 1. HAYSTACK SHAPE: the _s split serializes haystack_sessions as LongMemEvalTurn[][] (each inner array is one session's turns directly) plus a parallel `haystack_session_ids: string[]` field. The pre-v0.35.1.1 adapter expected only the oracle `{session_id, turns}` shape and crashed with `session.turns is undefined` on every question. Fix: new `normalizeSessions` helper accepts both shapes, mirroring the proven `normalizeSessions` in gbrain-evals/eval/runner/longmemeval.ts. 2. SLUG VALIDATOR: the _s split's session_ids look like `sharegpt_yywfIrx_0` — underscored and mixed-case. The v0.32.7 CJK wave's `validatePageSlug` rejects both (allowed set is `[a-z0-9-]` case-insensitive, slash-separated). Fix: `sanitizeSessionIdForSlug` lowercases and replaces `_` + `.` + any other non-[a-z0-9-] character with `-`. The frontmatter `session_id:` keeps the original verbatim for downstream JSONL emit; only the SLUG is rewritten. 3. INTERFACE: `LongMemEvalQuestion.haystack_sessions` typed as a union of `LongMemEvalSession[] | LongMemEvalTurn[][]` so TypeScript callers see both shapes are accepted. New `haystack_session_ids?: string[]` field documented as parallel to the array-of-turns shape. Pre-v0.35.1.1 caught by a fresh smoke pre-spend (3 questions × ZE @ 2560 → 3 errors). Post-fix: 3/3 OK with non-empty hypotheses, single-session recall measured (low on a 3-question sample but the pipeline runs). 2 new regression test cases pinning: - _s split shape normalizes (slugs sanitized + frontmatter preserves original session_id + dates flow through) - _s split with missing haystack_session_ids synthesizes `lme_<question_id>_<i>` ids Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(cli): configure AI gateway before running gbrain eval longmemeval v0.28.8 skipped connectEngine() for `gbrain eval longmemeval` so the subcommand could run on machines without a configured brain. Side effect (silent until v0.35.1.0 made it observable via the embedder shootout): the gateway was never configureGateway()'d either, so the first embed call inside importFromContent crashed with "AI gateway is not configured. Call configureGateway() during engine connect." Fix: call configureGateway() before runEvalLongMemEval, mirroring the connectEngine() path. Reads `~/.gbrain/config.json` when present; falls back to env vars (GBRAIN_EMBEDDING_MODEL, GBRAIN_EMBEDDING_DIMENSIONS, OPENAI_API_KEY, etc.) when there's no config — preserving the v0.28.8 "runs on fresh machine" property. Gated on the --help short-circuit so `gbrain eval longmemeval --help` still works without spinning up the gateway. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore: v0.35.1.1 Bumps VERSION + package.json + CHANGELOG entry for the longmemeval fix wave. Three commits this branch: 1. fix(eval): adapter handles _s split + sanitizes session_id slugs 2. fix(cli): configure AI gateway before running gbrain eval longmemeval 3. chore: v0.35.1.1 Each commit independently bisects; CHANGELOG entry is the user-facing rollup. No schema migration; no breaking change. Caught pre-spend by smoking Phase 1 of the embedder shootout — would otherwise have wasted ~$476 in judge tokens across 7 cells. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * ci: retrigger workflows --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>