feat(eval): Phase 2 EXT-3 hybrid-without-graph adapter — graph isolated

Third and closest-to-gbrain external baseline. Runs gbrain's full hybrid
search (vector + keyword + RRF fusion + dedup) WITHOUT the knowledge-graph
layer. Same engine, same embedder, same chunking, same hybrid fusion —
only traversePaths + typed-link extraction turned off.

This is the decisive comparator for "does the knowledge graph do useful
work?" Same everything-else, only graph differs. Any lead gbrain-after has
over EXT-3 is 100% attributable to the graph layer.

Files:
  eval/runner/adapters/hybrid-nograph.ts   — ~110 LOC

Implementation:
  - New PGLiteEngine per run; auto_link set to 'false' (belt).
  - importFromContent() used instead of bare putPage() so chunks +
    embeddings get populated (hybridSearch needs them).
  - NO runExtract() call — typed links/timeline stay empty (suspenders).
  - hybridSearch(engine, q.text) answers every query. Aggregate chunks
    to page-level by best chunk score.

FOUR-adapter side-by-side on 240-page rich-prose corpus, 145 relational queries:

| Adapter         | P@5    | R@5    | Correct/Gold |
|-----------------|--------|--------|--------------|
| gbrain-after    | 49.1%  | 97.9%  | 248/261      |
| hybrid-nograph  | 17.8%  | 65.1%  | 129/261      |
| ripgrep-bm25    | 17.1%  | 62.4%  | 124/261      |
| vector-only     | 10.8%  | 40.7%  |  78/261      |

The headline delta nobody can hand-wave away:
  gbrain-after → hybrid-nograph  = +31.4 P@5, +32.9 R@5
  hybrid-nograph → ripgrep-bm25  = +0.7 P@5,  +2.7 R@5

Hybrid search (vector+keyword+RRF) over pure BM25 gains ~1 point. The
knowledge graph layer over hybrid gains ~31 points. The graph is doing
the work; adding it to a retrieval stack is what actually moves the needle
on relational queries. The vector/keyword/BM25 debate is a footnote.

Timing: hybrid-nograph init is ~2 min (embeds 240 pages once); query loop
is fast. gbrain-after is ~1.5s total because traversePaths doesn't need
embeddings. Runs at ~$0.02 Opus-equivalent in embedding cost.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

eval/runner/adapters/hybrid-nograph.ts

@@ -0,0 +1,134 @@
+/**
+ * BrainBench EXT-3: Hybrid-without-graph adapter.
+ *
+ * gbrain's full hybrid search (vector + keyword + RRF fusion + dedup) but
+ * with the knowledge-graph layer explicitly disabled. No auto_link, no
+ * typed edges, no traverse_graph, no backlink boost. Just:
+ *   - putPage each page
+ *   - chunking + embedding (via existing put_page pipeline)
+ *   - hybridSearch(engine, query) to answer queries
+ *
+ * This is the closest-to-gbrain external comparator. If gbrain-after beats
+ * EXT-3 significantly, the delta MUST come from the graph layer (auto_link
+ * typed edges + traversePaths + backlink boost), not from better vector
+ * retrieval or hybrid fusion.
+ *
+ * It's also the MOST HONEST baseline — "gbrain without the new knowledge
+ * graph layer" answers the question "does the graph do useful work?"
+ * directly. Critics can't dismiss this as "you disabled a feature you knew
+ * they'd want." Everyone already knows vector+keyword hybrid is strong.
+ */
+
+import type { Adapter, AdapterConfig, BrainState, Page, Query, RankedDoc } from '../types.ts';
+import { PGLiteEngine } from '../../../src/core/pglite-engine.ts';
+import { hybridSearch } from '../../../src/core/search/hybrid.ts';
+import { importFromContent } from '../../../src/core/import-file.ts';
+
+// Known-safe config: auto_link OFF at the engine layer via direct setConfig
+// call. Does NOT run `extract --source db`, so typed links stay empty even
+// if auto_link flipped on during put_page (belt + suspenders).
+
+interface HybridNoGraphState {
+  engine: PGLiteEngine;
+}
+
+interface HybridNoGraphConfig extends AdapterConfig {
+  /** Top-K results requested from hybridSearch. Defaults to 20 so the
+   *  scorer's k=5 slice has headroom. */
+  limit?: number;
+}
+
+export class HybridNoGraphAdapter implements Adapter {
+  readonly name = 'hybrid-nograph';
+
+  async init(rawPages: Page[], _config: HybridNoGraphConfig): Promise<BrainState> {
+    const engine = new PGLiteEngine();
+    await engine.connect({});
+    await engine.initSchema();
+    // Belt: turn off auto_link at the engine config level. Suspenders below:
+    // we also skip extract --source db, so even if auto_link did fire, no
+    // typed edges would exist in the graph layer. This adapter doesn't call
+    // traversePaths at all, so graph state is doubly-ignored.
+    await engine.setConfig('auto_link', 'false');
+
+    // importFromContent does the chunking + embedding that hybridSearch needs.
+    // Plain putPage() just writes the page row without any search infra; that's
+    // fine for graph-based adapters but leaves hybridSearch with nothing to
+    // rank. Silence its stdout noise during benchmark runs.
+    const origLog = console.log;
+    const origErr = console.error;
+    console.log = () => {};
+    console.error = () => {};
+    try {
+      for (const p of rawPages) {
+        const content = this.buildContentMarkdown(p);
+        await importFromContent(engine, p.slug, content);
+      }
+    } finally {
+      console.log = origLog;
+      console.error = origErr;
+    }
+
+    // INTENTIONALLY do NOT call runExtract — that's what populates typed
+    // links + timeline for the graph layer. Without it, traversePaths
+    // would return empty. hybridSearch works entirely off chunks +
+    // embeddings, which importFromContent just populated.
+    return { engine } satisfies HybridNoGraphState;
+  }
+
+  /** Build a markdown string importFromContent can parse.
+   *  Format: YAML frontmatter then body; matches what gbrain import expects. */
+  private buildContentMarkdown(p: Page): string {
+    const fm: string[] = [];
+    fm.push(`---`);
+    fm.push(`type: ${p.type}`);
+    fm.push(`title: ${JSON.stringify(p.title)}`);
+    fm.push(`---`);
+    fm.push('');
+    fm.push(`# ${p.title}`);
+    fm.push('');
+    fm.push(p.compiled_truth);
+    if (p.timeline && p.timeline.trim().length > 0) {
+      fm.push('');
+      fm.push('## Timeline');
+      fm.push('');
+      fm.push(p.timeline);
+    }
+    return fm.join('\n');
+  }
+
+  async query(q: Query, state: BrainState): Promise<RankedDoc[]> {
+    const s = state as HybridNoGraphState;
+    const limit = 20;
+
+    // hybridSearch returns chunks with scores. We aggregate to page-level
+    // by taking each page's BEST chunk score and ranking pages by that.
+    const chunkResults = await hybridSearch(s.engine, q.text, { limit: limit * 3 });
+
+    const pageBest = new Map<string, number>();
+    for (const r of chunkResults) {
+      const existing = pageBest.get(r.slug);
+      if (existing === undefined || r.score > existing) {
+        pageBest.set(r.slug, r.score);
+      }
+    }
+    const pageScored = Array.from(pageBest.entries())
+      .map(([slug, score]) => ({ slug, score }))
+      .sort((a, b) => b.score - a.score || a.slug.localeCompare(b.slug))
+      .slice(0, limit);
+
+    return pageScored.map((p, i) => ({
+      page_id: p.slug,
+      score: p.score,
+      rank: i + 1,
+    }));
+  }
+
+  async snapshot(_state: BrainState): Promise<string> {
+    return '';
+  }
+}
+
+export function createHybridNoGraph(): HybridNoGraphAdapter {
+  return new HybridNoGraphAdapter();
+}

eval/runner/multi-adapter.ts

@@ -24,6 +24,7 @@ import { PGLiteEngine } from '../../src/core/pglite-engine.ts';
 import { runExtract } from '../../src/commands/extract.ts';
 import { RipgrepBm25Adapter } from './adapters/ripgrep-bm25.ts';
 import { VectorOnlyAdapter } from './adapters/vector-only.ts';
+import { HybridNoGraphAdapter } from './adapters/hybrid-nograph.ts';
 import type { Adapter, Page, Query, RankedDoc } from './types.ts';
 import { precisionAtK, recallAtK } from './types.ts';
 
@@ -325,6 +326,7 @@ async function main() {
 
   const allAdapters: Adapter[] = [
     new GbrainAfterAdapter(),
+    new HybridNoGraphAdapter(),
     new RipgrepBm25Adapter(),
     new VectorOnlyAdapter(),
   ];