Chunk

.env.example

@@ -149,6 +149,11 @@ SEMANTIC_EXPANSION_CACHE_TTL=3600
 # HYBRID_RECENCY_WEIGHT=0.1
 # RERANK_EXPAND=1
 
+# Elbow detection filter: adaptive threshold based on score distribution (Kneedle algorithm)
+# Filters out low-relevance results by detecting the "elbow" point in the score curve
+# Improves precision by only returning results above the natural relevance drop-off
+# HYBRID_ELBOW_FILTER=0
+
 # Caching (embeddings and search results)
 # MAX_EMBED_CACHE=16384
 # HYBRID_RESULTS_CACHE=128

.github/workflows/ci.yml

@@ -65,14 +65,18 @@ jobs:
         python -c "from fastembed import TextEmbedding; m = TextEmbedding(model_name='BAAI/bge-base-en-v1.5'); list(m.embed(['test']))"
 
     - name: Run tests
-      run: pytest -q
-      
+      run: pytest -q --junitxml=test-results.xml
+
     - name: Upload test results
       uses: actions/upload-artifact@v4
       if: always()
       with:
         name: test-results
-        path: |
-          .pytest_cache/
-          test-results.xml
+        path: test-results.xml
         retention-days: 7
+
+    - name: Test Summary
+      uses: test-summary/action@v2
+      if: always()
+      with:
+        paths: test-results.xml

ctx-mcp-bridge/package.json

@@ -1,14 +1,15 @@
 {
   "name": "@context-engine-bridge/context-engine-mcp-bridge",
-  "version": "0.0.15",
+  "version": "0.0.16",
   "description": "Context Engine MCP bridge (http/stdio proxy combining indexer + memory servers)",
   "bin": {
     "ctxce": "bin/ctxce.js",
     "ctxce-bridge": "bin/ctxce.js"
   },
   "type": "module",
   "scripts": {
-    "start": "node bin/ctxce.js"
+    "start": "node bin/ctxce.js",
+    "postinstall": "node -e \"try{require('fs').chmodSync('bin/ctxce.js',0o755)}catch(e){}\""
   },
   "dependencies": {
     "@modelcontextprotocol/sdk": "^1.24.3",
@@ -20,4 +21,4 @@
   "engines": {
     "node": ">=18.0.0"
   }
-}
+}

deploy/kubernetes/configmap.yaml

@@ -151,3 +151,4 @@ data:
   USE_GPU_DECODER: '0'
   USE_TREE_SITTER: '1'
   WATCH_DEBOUNCE_SECS: '4'
+  PSEUDO_DEFER_TO_WORKER: '1'

docker-compose.yml

@@ -453,9 +453,12 @@ services:
       - LEX_SPARSE_NAME=${LEX_SPARSE_NAME:-}
       # Pattern vectors for structural code similarity
       - PATTERN_VECTORS=${PATTERN_VECTORS:-}
-      # Graph edges for symbol relationships
-      - INDEX_GRAPH_EDGES=${INDEX_GRAPH_EDGES:-1}
+      # Graph edges for symbol relationships (always on)
       - INDEX_GRAPH_EDGES_MODE=${INDEX_GRAPH_EDGES_MODE:-symbol}
+      # Defer pseudo-tag generation to watcher worker for faster initial indexing
+      - PSEUDO_DEFER_TO_WORKER=${PSEUDO_DEFER_TO_WORKER:-1}
+      # Parallel indexing - number of worker threads (default: 4, use -1 for CPU count)
+      - INDEX_WORKERS=${INDEX_WORKERS:-4}
     volumes:
       - workspace_pvc:/work:rw
       - codebase_pvc:/work/.codebase:rw
@@ -514,12 +517,13 @@ services:
       - LEX_SPARSE_NAME=${LEX_SPARSE_NAME:-}
       # Pattern vectors for structural code similarity
       - PATTERN_VECTORS=${PATTERN_VECTORS:-}
-      # Graph edges for symbol relationships
-      - INDEX_GRAPH_EDGES=${INDEX_GRAPH_EDGES:-1}
+      # Graph edges for symbol relationships (always on - Qdrant flat graph)
       - INDEX_GRAPH_EDGES_MODE=${INDEX_GRAPH_EDGES_MODE:-symbol}
       - GRAPH_BACKFILL_ENABLED=${GRAPH_BACKFILL_ENABLED:-1}
-      # Neo4j graph backend (when set, edges go to Neo4j instead of Qdrant _graph collection)
+      # Neo4j graph backend (optional - takes precedence over Qdrant flat graph)
       - NEO4J_GRAPH=${NEO4J_GRAPH:-}
+      # Defer pseudo-tag generation - watcher runs backfill worker thread
+      - PSEUDO_DEFER_TO_WORKER=${PSEUDO_DEFER_TO_WORKER:-1}
     volumes:
       - workspace_pvc:/work:rw
       - codebase_pvc:/work/.codebase:rw

docs/CONFIGURATION.md

@@ -377,12 +377,26 @@ REFRAG_RUNTIME=glm  # or openai, minimax, llamacpp
 
 ### Pseudo Backfill Worker
 
-Deferred pseudo/tag generation runs asynchronously after initial indexing.
+Deferred pseudo/tag generation runs asynchronously after initial indexing. This significantly speeds up initial indexing by skipping LLM-based pseudo-tag generation during the indexer run, deferring it to a background worker thread in the watcher service.
 
 | Name | Description | Default |
 |------|-------------|---------|
 | PSEUDO_BACKFILL_ENABLED | Enable async pseudo/tag backfill worker | 0 (disabled) |
-| PSEUDO_DEFER_TO_WORKER | Skip inline pseudo, defer to backfill worker | 0 (disabled) |
+| PSEUDO_DEFER_TO_WORKER | Skip inline pseudo, defer to backfill worker | 1 (enabled) |
+| GRAPH_BACKFILL_ENABLED | Enable graph edge backfill in watcher worker | 1 (enabled) |
+
+**How it works:**
+1. When `PSEUDO_DEFER_TO_WORKER=1`, the indexer generates only base chunks (no pseudo-tags)
+2. The watcher service starts a `_start_pseudo_backfill_worker` daemon thread
+3. This thread periodically calls `pseudo_backfill_tick()` to enrich chunks with LLM-generated tags
+4. If `GRAPH_BACKFILL_ENABLED=1`, it also calls `graph_backfill_tick()` to populate symbol graph edges
+
+**Benefits:**
+- Initial indexing is 2-5x faster (no LLM calls blocking indexer)
+- Background enrichment happens continuously without blocking searches
+- Failed LLM calls don't break indexing; worker retries automatically
+
+**Recommended for production:** Enable both for fastest initial indexing with eventual enrichment.
 
 ### Adaptive Span Sizing
 
@@ -523,6 +537,7 @@ Useful for Kubernetes deployments where a shared filesystem is not reliable.
 | CODEBASE_STATE_REDIS_LOCK_WAIT_MS | Redis lock wait in ms | 2000 |
 | CODEBASE_STATE_REDIS_SOCKET_TIMEOUT | Redis socket timeout in seconds | 2 |
 | CODEBASE_STATE_REDIS_CONNECT_TIMEOUT | Redis connect timeout in seconds | 2 |
+| CODEBASE_STATE_REDIS_MAX_CONNECTIONS | Redis connection pool size limit | 10 |
 
 ### Semantic Expansion
 

pyproject.toml

@@ -58,6 +58,7 @@ dependencies = [
     "rich>=13.0.0",
     "typer>=0.9.0",
     "requests>=2.28.0",
+    "xxhash>=3.0.0",
 ]
 
 [project.optional-dependencies]