fix: full-pipeline eval + source overlap concept gate + 3 production bugs#29

Merged

7xuanlu merged 22 commits intomainfrom

feature/fullpipeline-eval

Apr 28, 2026

Commits on Apr 28, 2026

fix: full-pipeline eval with Batch API generation + flat cache reuse

7xuanlu
and
claude
committed
fix: use Qwen3.5-9B for enrichment in full-pipeline eval

7xuanlu
and
claude
committed
fix: correct 9B model ID (qwen3.5-9b not qwen35-9b)

7xuanlu
and
claude
committed
fix: extract_json_array misparses single KG objects with inner arrays

7xuanlu
and
claude
committed
fix: single-DB eval + enrichment_steps for concept distillation

7xuanlu
and
claude
committed
fix: batch size probe for on-device extraction ceiling

7xuanlu
and
claude
committed
fix: Batch API enrichment for eval (entity extraction via Haiku)

7xuanlu
and
claude
committed
fix: all LLM enrichment phases via Batch API (Haiku)

7xuanlu
and
claude
committed
fix: smoke test passes — title query + hallucination check fixed

7xuanlu
and
claude
committed
fix: persist enriched DB for resume (no more wasted API costs)

7xuanlu
and
claude
committed
fix: robust resume — detect partial vs complete enrichment

7xuanlu
and
claude
committed
fix: parallel batch submissions for extraction + title enrichment

7xuanlu
and
claude
committed
fix: parallel batch enrichment + CLI judge for Max plan

7xuanlu
and
claude
committed
fix: judge prompt consistency + revert concept cap + drop flat

7xuanlu
and
claude
committed
fix: task-specific judge prompts matching benchmark papers

7xuanlu
and
claude
committed
fix: CLI judge tests with configurable concurrency + LoCoMo CLI judge

7xuanlu
and
claude
committed
fix: return relevance scores from search_concepts (concept noise prep)

7xuanlu
and
claude
committed
fix: source overlap gate filters irrelevant concepts from context

7xuanlu
and
claude
committed
fix: remove dead params + add overlap gate unit tests (review fixes)

7xuanlu
and
claude
committed
fix: configurable concept_min_overlap in DistillationConfig

7xuanlu
and
claude
committed
fix(hooks): split pre-push (fast) from CI coverage (informational)
7xuanlu
committed
fix(eval): clippy literal-with-empty-format-string in probe test

7xuanlu
and
claude
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix: full-pipeline eval + source overlap concept gate + 3 production bugs#29

fix: full-pipeline eval + source overlap concept gate + 3 production bugs#29
7xuanlu merged 22 commits intomainfrom
feature/fullpipeline-eval

Commits on Apr 28, 2026

fix: full-pipeline eval with Batch API generation + flat cache reuse

fix: use Qwen3.5-9B for enrichment in full-pipeline eval

fix: correct 9B model ID (qwen3.5-9b not qwen35-9b)

fix: extract_json_array misparses single KG objects with inner arrays

fix: single-DB eval + enrichment_steps for concept distillation

fix: batch size probe for on-device extraction ceiling

fix: Batch API enrichment for eval (entity extraction via Haiku)

fix: all LLM enrichment phases via Batch API (Haiku)

fix: smoke test passes — title query + hallucination check fixed

fix: persist enriched DB for resume (no more wasted API costs)

fix: robust resume — detect partial vs complete enrichment

fix: parallel batch submissions for extraction + title enrichment

fix: parallel batch enrichment + CLI judge for Max plan

fix: judge prompt consistency + revert concept cap + drop flat

fix: task-specific judge prompts matching benchmark papers

fix: CLI judge tests with configurable concurrency + LoCoMo CLI judge

fix: return relevance scores from search_concepts (concept noise prep)

fix: source overlap gate filters irrelevant concepts from context

fix: remove dead params + add overlap gate unit tests (review fixes)

fix: configurable concept_min_overlap in DistillationConfig

fix(hooks): split pre-push (fast) from CI coverage (informational)

fix(eval): clippy literal-with-empty-format-string in probe test