[benchmarks] Retrieval benchmark suite — recall@K vs Pinecone, Weaviate, pgvector

scripts/bench/README.md

@@ -0,0 +1,27 @@
+# Engram Benchmark Suite
+
+Compares Engram retrieval against Pinecone, Weaviate, and pgvector.
+
+## Quick Start
+
+```bash
+# 1. Start Engram Web API
+cd /path/to/Engram
+uvicorn engram-web.api.main:app --port 8000
+
+# 2. Install deps
+pip install -r scripts/bench/requirements.txt
+
+# 3. Run benchmark
+python scripts/bench/run.py --engram-only --dataset nfcorpus
+python scripts/bench/run.py --all --output report.md
+```
+
+## Datasets
+
+Uses BEIR subsets (nfcorpus, fiqa, scidocs) by default.
+Falls back to synthetic data if BEIR is not installed.
+
+## Output
+
+Generates `bench_report.md` with recall@K and latency metrics.

scripts/bench/__init__.py

@@ -0,0 +1 @@
+# Engram Benchmark Suite

scripts/bench/requirements.txt

@@ -0,0 +1,13 @@
+# Engram Benchmark Suite - Dependencies
+# Core
+requests>=2.28
+numpy>=1.24
+beir>=2.0
+
+# Vector DB clients (install as needed)
+# pinecone-client>=3.0    # pip install pinecone-client
+# weaviate-client>=4.0    # pip install weaviate-client  
+# psycopg2-binary>=2.9   # pip install psycopg2-binary (for pgvector)
+
+# Engram (local)
+# Install engram from source: pip install -e /path/to/Engram