@zuvia-software-solutions/code-mapper 2.6.3 → 2.6.5

package/dist/core/embeddings/nl-embed-worker.js

@@ -9,7 +9,7 @@
 import { pipeline } from '@huggingface/transformers';
 const MODEL_ID = 'Xenova/bge-small-en-v1.5';
 async function main() {
-    const extractor = await pipeline('feature-extraction', MODEL_ID, { quantized: true });
+    const extractor = await pipeline('feature-extraction', MODEL_ID, { dtype: 'q8' });
     process.send({ type: 'ready' });
     // Process messages from parent
     process.on('message', async (msg) => {

package/dist/core/embeddings/nl-embedder.d.ts

@@ -30,7 +30,7 @@ interface NlDocument {
     source: string;
     text: string;
 }
-/** Build NL documents from a node */
+/** Build NL documents from a node — keyword-dense, minimal tokens */
 export declare function extractNlTexts(node: NodeForNl): NlDocument[];
 /**
  * Build NL embeddings for all eligible nodes in the database.

package/dist/core/embeddings/nl-embedder.js

@@ -27,7 +27,7 @@ export async function initNlEmbedder() {
         if (env.backends?.onnx?.wasm) {
             env.backends.onnx.wasm.numThreads = Math.max(1, cpuCount);
         }
-        extractor = await pipeline('feature-extraction', MODEL_ID, { quantized: true });
+        extractor = await pipeline('feature-extraction', MODEL_ID, { dtype: 'q8' });
     })();
     return loadPromise;
 }
@@ -147,11 +147,19 @@ function extractParamNames(content) {
         .map(p => expandIdentifier(p))
         .join(', ');
 }
-/** Build NL documents from a node */
+/** Strip noise tokens that waste tokenizer budget without adding semantic value */
+function condense(text) {
+    return text
+        .replace(/---[^-]*---/g, '') // section headers from comments
+        .replace(/[{}[\]()'",;:]/g, '') // punctuation
+        .replace(/\. /g, ' ') // sentence separators
+        .replace(/\s{2,}/g, ' ') // collapse whitespace
+        .trim();
+}
+/** Build NL documents from a node — keyword-dense, minimal tokens */
 export function extractNlTexts(node) {
     const docs = [];
-    const name = node.name;
-    const expandedName = expandIdentifier(name);
+    const expandedName = expandIdentifier(node.name);
     const dir = node.filePath.split('/').slice(-3, -1).join('/');
     // 1. Comment-based NL text (primary)
     const comment = extractFullComment(node.content);
@@ -159,22 +167,21 @@ export function extractNlTexts(node) {
         docs.push({
             nodeId: node.id,
             source: 'comment',
-            text: `${expandedName}: ${comment}. File: ${dir}`,
+            text: condense(`${expandedName} ${comment} ${dir}`),
         });
     }
-    // 2. Name + params + return type (always available)
+    // 2. Name + params (always available)
     const params = extractParamNames(node.content);
-    const parts = [expandedName];
-    if (params)
-        parts.push(`Parameters: ${params}`);
-    if (dir)
-        parts.push(`in ${dir}`);
     if (!comment) {
-        // Only add name-based doc if no comment (avoid duplication)
+        const parts = [expandedName];
+        if (params)
+            parts.push(params);
+        if (dir)
+            parts.push(dir);
         docs.push({
             nodeId: node.id,
             source: 'name',
-            text: parts.join('. '),
+            text: condense(parts.join(' ')),
         });
     }
     // 3. Enum/const values
@@ -184,7 +191,7 @@ export function extractNlTexts(node) {
             docs.push({
                 nodeId: node.id,
                 source: 'enum',
-                text: `${expandedName}: ${values}`,
+                text: condense(`${expandedName} ${values}`),
             });
         }
     }
@@ -274,8 +281,9 @@ export async function buildNlEmbeddings(db, onProgress) {
     // Find worker script path
     const thisDir = pathMod.dirname(fileURLToPath(import.meta.url));
     const workerScript = pathMod.join(thisDir, 'nl-embed-worker.js');
-    // Split work across workers
-    const ITEMS_PER_BATCH = 50;
+    // Split work across workers — larger batches reduce IPC round-trips
+    // and let the ONNX runtime amortize overhead across more items
+    const ITEMS_PER_BATCH = 256;
     let nextIdx = 0;
     let embedded = 0;
     const getNextBatch = () => {
@@ -432,3 +440,4 @@ export async function buildNlEmbeddings(db, onProgress) {
     }
     return { embedded, skipped, durationMs: Date.now() - t0 };
 }
+// touch

package/dist/core/incremental/refresh.js

@@ -311,6 +311,99 @@ export async function refreshFiles(db, repoPath, dirtyFiles) {
         }
     }
     // FTS5 auto-updates via triggers — no manual rebuild needed
+    // Phase 5: Rebuild graph-level analyses (communities, processes, interface dispatch)
+    // These are cheap (<300ms) but critical — stale communities/processes mislead agents.
+    // Load full graph from SQLite, re-run analyses, write results back.
+    try {
+        const { createKnowledgeGraph } = await import('../graph/graph.js');
+        const { processCommunities } = await import('../ingestion/community-processor.js');
+        const { processProcesses } = await import('../ingestion/process-processor.js');
+        const { insertNodesBatch, insertEdgesBatch } = await import('../db/adapter.js');
+        const { toNodeId, toEdgeId } = await import('../db/schema.js');
+        const graph = createKnowledgeGraph();
+        // Load all non-community/process nodes and edges into in-memory graph
+        const allNodes = db.prepare('SELECT * FROM nodes WHERE label NOT IN (\'Community\', \'Process\')').all();
+        const allEdges = db.prepare('SELECT * FROM edges WHERE type NOT IN (\'MEMBER_OF\', \'STEP_IN_PROCESS\')').all();
+        for (const row of allNodes) {
+            graph.addNode({
+                id: toNodeId(row.id),
+                label: row.label,
+                properties: {
+                    name: row.name ?? '', filePath: row.filePath ?? '',
+                    startLine: row.startLine ?? undefined, endLine: row.endLine ?? undefined,
+                    isExported: Boolean(row.isExported),
+                    description: row.description ?? undefined,
+                    parameterCount: row.parameterCount ?? undefined,
+                    returnType: row.returnType ?? undefined,
+                },
+            });
+        }
+        for (const row of allEdges) {
+            graph.addRelationship({
+                id: toEdgeId(row.id),
+                sourceId: toNodeId(row.sourceId),
+                targetId: toNodeId(row.targetId),
+                type: row.type,
+                confidence: row.confidence ?? 1.0,
+                reason: row.reason ?? '',
+            });
+        }
+        // Delete old community/process data from SQLite
+        db.exec('BEGIN');
+        db.prepare("DELETE FROM edges WHERE type IN ('MEMBER_OF', 'STEP_IN_PROCESS')").run();
+        db.prepare("DELETE FROM nodes WHERE label IN ('Community', 'Process')").run();
+        // Re-run community detection + process detection
+        const communityResult = await processCommunities(graph, () => { });
+        const processResult = await processProcesses(graph, communityResult.memberships, () => { });
+        // Write new community/process nodes + edges back to SQLite
+        const newNodes = [];
+        const newEdges = [];
+        for (const node of graph.iterNodes()) {
+            if (node.label === 'Community' || node.label === 'Process') {
+                newNodes.push({
+                    id: node.id,
+                    label: node.label,
+                    name: node.properties.name ?? '',
+                    filePath: node.properties.filePath ?? '',
+                    heuristicLabel: node.properties.heuristicLabel ?? null,
+                    cohesion: node.properties.cohesion ?? null,
+                    symbolCount: node.properties.symbolCount ?? null,
+                    keywords: Array.isArray(node.properties.keywords) ? node.properties.keywords.join(', ') : node.properties.keywords ?? null,
+                    processType: node.properties.processType ?? null,
+                    stepCount: node.properties.stepCount ?? null,
+                    communities: node.properties.communities ?? null,
+                    entryPointId: node.properties.entryPointId ?? null,
+                    terminalId: node.properties.terminalId ?? null,
+                });
+            }
+        }
+        for (const rel of graph.iterRelationships()) {
+            if (rel.type === 'MEMBER_OF' || rel.type === 'STEP_IN_PROCESS') {
+                newEdges.push({
+                    id: rel.id,
+                    sourceId: rel.sourceId,
+                    targetId: rel.targetId,
+                    type: rel.type,
+                    confidence: rel.confidence,
+                    reason: rel.reason,
+                    step: rel.step,
+                });
+            }
+        }
+        if (newNodes.length > 0)
+            insertNodesBatch(db, newNodes);
+        if (newEdges.length > 0)
+            insertEdgesBatch(db, newEdges);
+        db.exec('COMMIT');
+        console.error(`Code Mapper: refresh Phase 5 — ${communityResult.communities.length} communities, ${processResult.stats.totalProcesses} processes rebuilt`);
+    }
+    catch (err) {
+        try {
+            db.exec('ROLLBACK');
+        }
+        catch { }
+        console.error(`Code Mapper: Phase 5 graph rebuild failed: ${err instanceof Error ? err.message : err}`);
+    }
     return {
         filesProcessed: filesToProcess.length, filesSkipped,
         nodesDeleted, nodesInserted, edgesInserted,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@zuvia-software-solutions/code-mapper",
-  "version": "2.6.3",
+  "version": "2.6.5",
   "description": "Graph-powered code intelligence for AI agents. Index any codebase, query via MCP or CLI.",
   "author": "Abhigyan Patwari",
   "license": "PolyForm-Noncommercial-1.0.0",