Banchmark

.github/workflows/ci.yml

@@ -0,0 +1,58 @@
+# CI for the cgo C++ SIMD core and pure-Go fallback.
+#
+# The matrix runs natively on amd64 (portable scalar C++ path) and arm64
+# (NEON kernels), each with cgo on and off. Because the hosted runners
+# already cover both Linux architectures natively, the Makefile's
+# docker-verify target is intentionally NOT run here.
+name: CI
+
+on:
+  push:
+    branches: [main, banchmark]
+  pull_request:
+
+jobs:
+  test:
+    name: test (${{ matrix.os }}, cgo=${{ matrix.cgo }})
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, ubuntu-24.04-arm] # amd64 + arm64 (NEON)
+        cgo: [1, 0]
+    env:
+      CGO_ENABLED: ${{ matrix.cgo }}
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+
+      - name: go vet
+        run: go vet ./...
+
+      - name: go build
+        run: go build ./...
+
+      - name: go test
+        run: go test ./...
+
+  gofmt:
+    name: gofmt
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+
+      - name: gofmt
+        run: |
+          files=$(gofmt -l .)
+          if [ -n "$files" ]; then
+            echo "The following files are not gofmt-formatted:"
+            echo "$files"
+            exit 1
+          fi

CLAUDE.md

@@ -86,10 +86,6 @@ make clean
 **API (`internal/api/`)**
 - Gin HTTP server (`cmd/vectordb-server`)
 
-**Legacy (`db/`, `storage/`)**
-- Older brute-force engine (linear scan + top-k heap), kept as scaffold;
-  its Search uses the batched `CosineSimilarityMany` kernel
-
 ### Conventions for the SIMD core
 - Never change `pkg/vectormath` public signatures; callers must not care
   which kernel is active
@@ -101,7 +97,9 @@ make clean
   locally and via `make docker-verify`
 
 ### Known quirks
-- `config.yaml` contains absolute macOS paths; tests use `t.TempDir()`
+- `config.yaml` uses relative paths (`data/`, `logs/`) resolved against the
+  process working directory — run binaries from the repo root (or pass an
+  absolute `-config` path and adjust the paths); tests use `t.TempDir()`
 
 ### Configuration System
 - Main config file: `config.yaml` (server host/port, storage path, index

README.md

@@ -24,7 +24,6 @@ vectorDB/
 │       ├── scalar/        #   pure-Go kernels (CGO_ENABLED=0 + parity reference)
 │       └── simd/          #   C++17 SIMD core via cgo (NEON on aarch64)
 ├── persistence/           # BadgerDB store
-├── db/, storage/          # Legacy brute-force engine + storage interfaces
 ├── proto/                 # Protocol buffer definitions
 ├── docs/                  # SIMD benchmark report
 ├── test/                  # Integration tests
@@ -142,9 +141,9 @@ A Go→C call costs roughly 40–50 ns. That fact shaped the API:
   per-pair call has to be cheap by itself. Fusion is the mitigation there.
 - *Paths we control do batch*: `CosineSimilarityBatch`/`CosineSimilarityMany`
   score one query against N vectors in a **single** crossing (flattened
-  row-major buffer, query norm computed once). The brute-force scan in
-  `db/engine.go` uses this — 10,000 vectors cost one crossing instead of
-  10,000, which is the 561 µs → 114 µs row in the tables.
+  row-major buffer, query norm computed once) — 10,000 vectors cost one
+  crossing instead of 10,000, which is the 561 µs → 114 µs row in the
+  tables.
 - *Zero allocations*: the fused kernel reports zero-norm inputs via a NaN
   sentinel instead of an out-pointer, because any Go pointer passed to C
   escapes to the heap — the sentinel keeps the hot path allocation-free.
@@ -158,7 +157,7 @@ reassociates float32 additions.
 
 ### Key design points
 
-- **Frozen public API** — every caller (`internal/index`, `db`, tests) is
+- **Frozen public API** — every caller (`internal/index`, tests) is
   untouched; the kernel swap is invisible above `pkg/vectormath`.
 - **Pure-Go fallback** — `CGO_ENABLED=0` builds and passes the full test
   suite anywhere Go runs, no C++ toolchain required.
@@ -287,8 +286,23 @@ index:
 
 database:
   max_vectors: 1000000
+
+logging:
+  level: info
+  encoding: json
+  output_paths:
+    - stdout
+    - logs/vectordb.log
+  dev_mode: false
+
+badger:
+  path: data
 ```
 
+Paths (`storage.path`, `badger.path`, file entries in
+`logging.output_paths`) are relative to the process working directory;
+missing directories are created on startup.
+
 ## Development
 ```bash
 make build

config.yaml

@@ -3,7 +3,7 @@ server:
   port: 8080
 
 storage:
-  path: /Users/ishaanbajpai/Desktop/bitCamp/vectorDB/data
+  path: data
 
 index:
   type: hnsw
@@ -18,7 +18,7 @@ logging:
   output_paths:
     - stdout
     - logs/vectordb.log
-  dev_mode: false 
+  dev_mode: false
 
 badger:
-  path: /Users/ishaanbajpai/Desktop/bitCamp/vectorDB/data
+  path: data

docs/simd-benchmark-report.md

@@ -42,7 +42,8 @@ Notes:
 | pure Go fused, batched | 1093 µs | 109 ns |
 
 The batched kernel amortizes the cgo crossing to nothing and lets the NEON
-loop stream; it is used by the brute-force scan in `db/engine.go`.
+loop stream; it is exposed as the public
+`CosineSimilarityBatch`/`CosineSimilarityMany` API.
 
 ## End-to-end HNSW (dim=128, M=16, efConstruction=200; count=3, benchtime=1000x)
 

internal/config/config.go

@@ -75,5 +75,14 @@ func DefaultConfig() *Config {
 		Database: DatabaseConfig{
 			MaxVectors: 1000000,
 		},
+		Logging: logger.Config{
+			Level:       "info",
+			Encoding:    "json",
+			OutputPaths: []string{"stdout"},
+			DevMode:     false,
+		},
+		Badger: BadgerConfig{
+			Path: "data",
+		},
 	}
 }