fix: resolve all dogfood v2.2.0 bugs

generated/DOGFOOD_REPORT_v2.2.0.md

@@ -42,6 +42,8 @@
 
 **Fix:** Treat `.` (or current dir equivalent) as `null`/no filter in `structureData()`.
 
+> **FIXED** — `structureData()` now normalizes the directory argument and treats `"."` as null/no filter. (`src/structure.js`)
+
 ### 2. Stale embeddings after rebuild (Medium severity)
 
 - After an incremental `build`, embedding `node_id`s become orphaned (e.g. old IDs in 3077-range, new IDs in 4335-range)
@@ -52,6 +54,8 @@
 
 **Fix:** Either preserve node IDs across rebuilds, invalidate embeddings when node IDs change, or warn the user to re-run `embed`.
 
+> **FIXED** — Build now invalidates embeddings alongside nodes. Full builds clear the embeddings table entirely. Incremental builds delete embeddings for affected files before deleting their nodes (order matters — need node IDs to find them). After the build, any remaining orphaned embeddings trigger a warning: `"N embeddings are orphaned (nodes changed). Run codegraph embed to refresh."` (`src/builder.js`)
+
 ### 3. `embed` default model requires HuggingFace auth (Medium severity)
 
 - `codegraph embed .` crashes with `Error: Unauthorized access to file` for the default `jina-code` model
@@ -61,6 +65,8 @@
 
 **Fix:** Either default to a public model (e.g. `minilm`), auto-fallback to `minilm` on auth failure, or catch the error and provide a clear message with instructions.
 
+> **FIXED** — Default model changed from `nomic-v1.5` (gated, requires HF_TOKEN) to `minilm` (public, 23MB, always works). Additionally, `loadModel()` now catches auth/download failures and prints a clear message with options (set HF_TOKEN or use `--model minilm`) instead of crashing with a raw stack trace. (`src/embedder.js`, `src/cli.js`)
+
 ### 4. Cross-language false positive in export (Low severity)
 
 - One low-confidence (0.3) call edge: `main` (build.rs) → `setup` (tests/unit/structure.test.js)
@@ -69,6 +75,8 @@
 
 **Fix:** Export commands could support a `--min-confidence` filter, or the default export could exclude edges below a threshold (e.g. 0.5).
 
+> **FIXED** — Added `--min-confidence <score>` option to the `export` command (default: 0.5). All three formats (DOT, Mermaid, JSON) filter edges by confidence at the SQL level. The 0.3-confidence false positive is excluded by default. Users can pass `--min-confidence 0` to include all edges. (`src/export.js`, `src/cli.js`)
+
 ## `--no-tests` Flag
 
 Tested on `stats` and `map` — both correctly filter out test files:
@@ -80,3 +88,5 @@ Tested on `stats` and `map` — both correctly filter out test files:
 - `embed --model minilm` successfully generated 392 embeddings (384d)
 - `search "build graph"` returned 15 results after fresh embeddings (top hit: 37.9% `test_triangle_cycle`)
 - Search quality is reasonable but not ideal — `buildGraph` itself didn't appear in results for "build graph"
+
+> **FIXED** — Embedding text now includes a readable split of the identifier name (e.g. `buildGraph` → `"function buildGraph (build Graph) in src/builder.js"`). This lets the model naturally associate "build graph" queries with `buildGraph` without needing hybrid search. camelCase, PascalCase, snake_case, and kebab-case are all handled. (`src/embedder.js`)

src/builder.js

@@ -402,13 +402,30 @@ export async function buildGraph(rootDir, opts = {}) {
     return;
   }
 
+  // Check if embeddings table exists (created by `embed`, not by initSchema)
+  let hasEmbeddings = false;
+  try {
+    db.prepare('SELECT 1 FROM embeddings LIMIT 1').get();
+    hasEmbeddings = true;
+  } catch {
+    /* table doesn't exist */
+  }
+
   if (isFullBuild) {
+    const deletions =
+      'PRAGMA foreign_keys = OFF; DELETE FROM node_metrics; DELETE FROM edges; DELETE FROM nodes; PRAGMA foreign_keys = ON;';
     db.exec(
-      'PRAGMA foreign_keys = OFF; DELETE FROM node_metrics; DELETE FROM edges; DELETE FROM nodes; PRAGMA foreign_keys = ON;',
+      hasEmbeddings
+        ? `${deletions.replace('PRAGMA foreign_keys = ON;', '')} DELETE FROM embeddings; PRAGMA foreign_keys = ON;`
+        : deletions,
     );
   } else {
     info(`Incremental: ${parseChanges.length} changed, ${removed.length} removed`);
-    // Remove metrics/edges/nodes for changed and removed files
+    // Remove embeddings/metrics/edges/nodes for changed and removed files
+    // Embeddings must be deleted BEFORE nodes (we need node IDs to find them)
+    const deleteEmbeddingsForFile = hasEmbeddings
+      ? db.prepare('DELETE FROM embeddings WHERE node_id IN (SELECT id FROM nodes WHERE file = ?)')
+      : null;
     const deleteNodesForFile = db.prepare('DELETE FROM nodes WHERE file = ?');
     const deleteEdgesForFile = db.prepare(`
       DELETE FROM edges WHERE source_id IN (SELECT id FROM nodes WHERE file = @f)
@@ -418,12 +435,14 @@ export async function buildGraph(rootDir, opts = {}) {
       'DELETE FROM node_metrics WHERE node_id IN (SELECT id FROM nodes WHERE file = ?)',
     );
     for (const relPath of removed) {
+      deleteEmbeddingsForFile?.run(relPath);
       deleteEdgesForFile.run({ f: relPath });
       deleteMetricsForFile.run(relPath);
       deleteNodesForFile.run(relPath);
     }
     for (const item of parseChanges) {
       const relPath = item.relPath || normalizePath(path.relative(rootDir, item.file));
+      deleteEmbeddingsForFile?.run(relPath);
       deleteEdgesForFile.run({ f: relPath });
       deleteMetricsForFile.run(relPath);
       deleteNodesForFile.run(relPath);
@@ -823,6 +842,23 @@ export async function buildGraph(rootDir, opts = {}) {
   const nodeCount = db.prepare('SELECT COUNT(*) as c FROM nodes').get().c;
   info(`Graph built: ${nodeCount} nodes, ${edgeCount} edges`);
   info(`Stored in ${dbPath}`);
+
+  // Warn about orphaned embeddings that no longer match any node
+  if (hasEmbeddings) {
+    try {
+      const orphaned = db
+        .prepare('SELECT COUNT(*) as c FROM embeddings WHERE node_id NOT IN (SELECT id FROM nodes)')
+        .get().c;
+      if (orphaned > 0) {
+        warn(
+          `${orphaned} embeddings are orphaned (nodes changed). Run "codegraph embed" to refresh.`,
+        );
+      }
+    } catch {
+      /* ignore — embeddings table may have been dropped */
+    }
+  }
+
   db.close();
 
   // Write journal header after successful build

src/cli.js

@@ -272,10 +272,15 @@ program
   .option('--functions', 'Function-level graph instead of file-level')
   .option('-T, --no-tests', 'Exclude test/spec files')
   .option('--include-tests', 'Include test/spec files (overrides excludeTests config)')
+  .option('--min-confidence <score>', 'Minimum edge confidence threshold (default: 0.5)', '0.5')
   .option('-o, --output <file>', 'Write to file instead of stdout')
   .action((opts) => {
     const db = new Database(findDbPath(opts.db), { readonly: true });
-    const exportOpts = { fileLevel: !opts.functions, noTests: resolveNoTests(opts) };
+    const exportOpts = {
+      fileLevel: !opts.functions,
+      noTests: resolveNoTests(opts),
+      minConfidence: parseFloat(opts.minConfidence),
+    };
 
     let output;
     switch (opts.format) {
@@ -412,7 +417,7 @@ program
   .action(() => {
     console.log('\nAvailable embedding models:\n');
     for (const [key, config] of Object.entries(MODELS)) {
-      const def = key === 'nomic-v1.5' ? ' (default)' : '';
+      const def = key === 'minilm' ? ' (default)' : '';
       console.log(`  ${key.padEnd(12)} ${String(config.dim).padStart(4)}d  ${config.desc}${def}`);
     }
     console.log('\nUsage: codegraph embed --model <name>');
@@ -426,8 +431,8 @@ program
   )
   .option(
     '-m, --model <name>',
-    'Embedding model: minilm, jina-small, jina-base, jina-code, nomic, nomic-v1.5 (default), bge-large. Run `codegraph models` for details',
-    'nomic-v1.5',
+    'Embedding model: minilm (default), jina-small, jina-base, jina-code, nomic, nomic-v1.5, bge-large. Run `codegraph models` for details',
+    'minilm',
   )
   .action(async (dir, opts) => {
     const root = path.resolve(dir || '.');

src/embedder.js

@@ -4,6 +4,18 @@ import Database from 'better-sqlite3';
 import { findDbPath, openReadonlyOrFail } from './db.js';
 import { warn } from './logger.js';
 
+/**
+ * Split an identifier into readable words.
+ * camelCase/PascalCase → "camel Case", snake_case → "snake case", kebab-case → "kebab case"
+ */
+function splitIdentifier(name) {
+  return name
+    .replace(/([a-z])([A-Z])/g, '$1 $2')
+    .replace(/([A-Z]+)([A-Z][a-z])/g, '$1 $2')
+    .replace(/[_-]+/g, ' ')
+    .trim();
+}
+
 // Lazy-load transformers (heavy, optional module)
 let pipeline = null;
 let _cos_sim = null;
@@ -55,7 +67,7 @@ export const MODELS = {
   },
 };
 
-export const DEFAULT_MODEL = 'nomic-v1.5';
+export const DEFAULT_MODEL = 'minilm';
 const BATCH_SIZE_MAP = {
   minilm: 32,
   'jina-small': 16,
@@ -103,8 +115,27 @@ async function loadModel(modelKey) {
   _cos_sim = transformers.cos_sim;
 
   console.log(`Loading embedding model: ${config.name} (${config.dim}d)...`);
-  const opts = config.quantized ? { quantized: true } : {};
-  extractor = await pipeline('feature-extraction', config.name, opts);
+  const pipelineOpts = config.quantized ? { quantized: true } : {};
+  try {
+    extractor = await pipeline('feature-extraction', config.name, pipelineOpts);
+  } catch (err) {
+    const msg = err.message || String(err);
+    if (msg.includes('Unauthorized') || msg.includes('401') || msg.includes('gated')) {
+      console.error(
+        `\nModel "${config.name}" requires authentication.\n` +
+          `This model is gated on HuggingFace and needs an access token.\n\n` +
+          `Options:\n` +
+          `  1. Set HF_TOKEN env var: export HF_TOKEN=hf_...\n` +
+          `  2. Use a public model instead: codegraph embed --model minilm\n`,
+      );
+    } else {
+      console.error(
+        `\nFailed to load model "${config.name}": ${msg}\n` +
+          `Try a different model: codegraph embed --model minilm\n`,
+      );
+    }
+    process.exit(1);
+  }
   activeModel = config.name;
   console.log('Model loaded.');
   return { extractor, config };
@@ -219,7 +250,8 @@ export async function buildEmbeddings(rootDir, modelKey, customDbPath) {
         : Math.min(lines.length, startLine + 15);
       const context = lines.slice(startLine, endLine).join('\n');
 
-      const text = `${node.kind} ${node.name} in ${file}\n${context}`;
+      const readable = splitIdentifier(node.name);
+      const text = `${node.kind} ${node.name} (${readable}) in ${file}\n${context}`;
       texts.push(text);
       nodeIds.push(node.id);
       previews.push(`${node.name} (${node.kind}) -- ${file}:${node.line}`);

src/export.js

@@ -1,12 +1,15 @@
 import path from 'node:path';
 import { isTestFile } from './queries.js';
 
+const DEFAULT_MIN_CONFIDENCE = 0.5;
+
 /**
  * Export the dependency graph in DOT (Graphviz) format.
  */
 export function exportDOT(db, opts = {}) {
   const fileLevel = opts.fileLevel !== false;
   const noTests = opts.noTests || false;
+  const minConf = opts.minConfidence ?? DEFAULT_MIN_CONFIDENCE;
   const lines = [
     'digraph codegraph {',
     '  rankdir=LR;',
@@ -23,8 +26,9 @@ export function exportDOT(db, opts = {}) {
       JOIN nodes n1 ON e.source_id = n1.id
       JOIN nodes n2 ON e.target_id = n2.id
       WHERE n1.file != n2.file AND e.kind IN ('imports', 'imports-type', 'calls')
+        AND e.confidence >= ?
     `)
-      .all();
+      .all(minConf);
     if (noTests) edges = edges.filter((e) => !isTestFile(e.source) && !isTestFile(e.target));
 
     // Try to use directory nodes from DB (built by structure analysis)
@@ -102,8 +106,9 @@ export function exportDOT(db, opts = {}) {
       JOIN nodes n2 ON e.target_id = n2.id
       WHERE n1.kind IN ('function', 'method', 'class', 'interface', 'type', 'struct', 'enum', 'trait', 'record', 'module') AND n2.kind IN ('function', 'method', 'class', 'interface', 'type', 'struct', 'enum', 'trait', 'record', 'module')
       AND e.kind = 'calls'
+      AND e.confidence >= ?
     `)
-      .all();
+      .all(minConf);
     if (noTests)
       edges = edges.filter((e) => !isTestFile(e.source_file) && !isTestFile(e.target_file));
 
@@ -126,6 +131,7 @@ export function exportDOT(db, opts = {}) {
 export function exportMermaid(db, opts = {}) {
   const fileLevel = opts.fileLevel !== false;
   const noTests = opts.noTests || false;
+  const minConf = opts.minConfidence ?? DEFAULT_MIN_CONFIDENCE;
   const lines = ['graph LR'];
 
   if (fileLevel) {
@@ -136,8 +142,9 @@ export function exportMermaid(db, opts = {}) {
       JOIN nodes n1 ON e.source_id = n1.id
       JOIN nodes n2 ON e.target_id = n2.id
       WHERE n1.file != n2.file AND e.kind IN ('imports', 'imports-type', 'calls')
+        AND e.confidence >= ?
     `)
-      .all();
+      .all(minConf);
     if (noTests) edges = edges.filter((e) => !isTestFile(e.source) && !isTestFile(e.target));
 
     for (const { source, target } of edges) {
@@ -155,8 +162,9 @@ export function exportMermaid(db, opts = {}) {
       JOIN nodes n2 ON e.target_id = n2.id
       WHERE n1.kind IN ('function', 'method', 'class', 'interface', 'type', 'struct', 'enum', 'trait', 'record', 'module') AND n2.kind IN ('function', 'method', 'class', 'interface', 'type', 'struct', 'enum', 'trait', 'record', 'module')
       AND e.kind = 'calls'
+      AND e.confidence >= ?
     `)
-      .all();
+      .all(minConf);
     if (noTests)
       edges = edges.filter((e) => !isTestFile(e.source_file) && !isTestFile(e.target_file));
 
@@ -175,6 +183,7 @@ export function exportMermaid(db, opts = {}) {
  */
 export function exportJSON(db, opts = {}) {
   const noTests = opts.noTests || false;
+  const minConf = opts.minConfidence ?? DEFAULT_MIN_CONFIDENCE;
 
   let nodes = db
     .prepare(`
@@ -185,13 +194,13 @@ export function exportJSON(db, opts = {}) {
 
   let edges = db
     .prepare(`
-    SELECT DISTINCT n1.file AS source, n2.file AS target, e.kind
+    SELECT DISTINCT n1.file AS source, n2.file AS target, e.kind, e.confidence
     FROM edges e
     JOIN nodes n1 ON e.source_id = n1.id
     JOIN nodes n2 ON e.target_id = n2.id
-    WHERE n1.file != n2.file
+    WHERE n1.file != n2.file AND e.confidence >= ?
   `)
-    .all();
+    .all(minConf);
   if (noTests) edges = edges.filter((e) => !isTestFile(e.source) && !isTestFile(e.target));
 
   return { nodes, edges };

src/structure.js

@@ -231,7 +231,8 @@ export function buildStructure(db, fileSymbols, _rootDir, lineCountMap, director
  */
 export function structureData(customDbPath, opts = {}) {
   const db = openReadonlyOrFail(customDbPath);
-  const filterDir = opts.directory || null;
+  const rawDir = opts.directory || null;
+  const filterDir = rawDir && normalizePath(rawDir) !== '.' ? rawDir : null;
   const maxDepth = opts.depth || null;
   const sortBy = opts.sort || 'files';
   const noTests = opts.noTests || false;