embed cache — LRU at /v1/embed for repeat-query elimination

Adds CachedProvider wrapping the embedding Provider with a thread-safe
LRU keyed on (effective_model, sha256(text)) → []float32. Repeat
queries return the stored vector without round-tripping to Ollama.

Why this matters: the staffing 500K test (memory project_golang_lakehouse)
documented that the staffing co-pilot replays many of the same query
texts ("forklift driver IL", "welder Chicago", "warehouse safety", etc).
Each repeat paid the ~50ms Ollama round-trip. Cached repeats now serve
in <1µs (LRU lookup + sha256 of input).

Memory budget: ~3 KiB per entry at d=768. Default 10K entries ≈ 30 MiB.
Configurable via [embedd].cache_size; 0 disables (pass-through mode).

Per-text caching, not per-batch — a batch with mixed hits/misses only
fetches the misses upstream, then merges the result preserving caller
input order. Three-text batch with one miss = one upstream call for
that one text instead of three.

Implementation:
  internal/embed/cached.go (NEW, 150 LoC)
    CachedProvider implements Provider; uses hashicorp/golang-lru/v2.
    Key shape: "<model>:<sha256-hex>". Empty model resolves to
    defaultModel (request-derived) for the key — NOT res.Model
    (upstream-derived), so future requests with same input shape
    hit the same key. Caught by TestCachedProvider_EmptyModelResolvesToDefault.
    Atomic hit/miss counters + Stats() + HitRate() + Len().

  internal/embed/cached_test.go (NEW, 12 test funcs)
    Pass-through-when-zero, hit-on-repeat, mixed-batch only fetches
    misses, model-key isolation, empty-model resolves to default,
    LRU eviction at cap, error propagation, all-hits synthesized
    without upstream call, hit-rate accumulation, empty-texts
    rejected, concurrent-safe (50 goroutines × 100 calls), key
    stability + distinctness.

  internal/shared/config.go
    EmbeddConfig.CacheSize (toml: cache_size). Default 10000.

  cmd/embedd/main.go
    Wraps Ollama Provider with CachedProvider on startup. Adds
    /embed/stats endpoint exposing hits / misses / hit_rate / size.
    Operators check the rate to confirm the cache is working
    (high rate = good) or sized wrong (low rate + many misses on a
    workload that should have repeats).

  cmd/embedd/main_test.go
    Stats endpoint tests — disabled mode shape, enabled mode tracks
    hits + misses across repeat calls.

One real bug caught by my own test:
  Initial implementation cached under res.Model (upstream-resolved)
  rather than effectiveModel (request-resolved). A request with
  model="" caching under "test-model" (Ollama's default), then a
  request with model="the-default" (our config default) missing
  the cache. Fix: always use the request-derived effectiveModel
  for keys; that's the predictable side. Locked by
  TestCachedProvider_EmptyModelResolvesToDefault.

Verified:
  go test -count=1 ./internal/embed/  — all 12 cached tests + 6 ollama tests green
  go test -count=1 ./cmd/embedd/      — stats endpoint tests green
  just verify                          — vet + test + 9 smokes 33s

Production benefit:
  ~50ms Ollama round-trip → <1µs cache lookup for cached entries.
  At 10K-entry default + ~30% repeat rate (typical staffing co-pilot
  workload), saves several seconds per staffer-query session.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

cmd/embedd/main.go

@@ -41,9 +41,20 @@ func main() {
 		os.Exit(1)
 	}
 
-	h := &handlers{
+	// Wrap the upstream provider in an LRU cache so repeat queries
-		provider: embed.NewOllama(cfg.Embedd.ProviderURL, cfg.Embedd.DefaultModel),
+	// (the staffing co-pilot replays many of the same texts) bypass
+	// the ~50ms Ollama round-trip. Cache size 0 = pass-through.
+	base := embed.NewOllama(cfg.Embedd.ProviderURL, cfg.Embedd.DefaultModel)
+	cached, err := embed.NewCachedProvider(base, cfg.Embedd.DefaultModel, cfg.Embedd.CacheSize)
+	if err != nil {
+		slog.Error("embed cache", "err", err)
+		os.Exit(1)
 	}
+	slog.Info("embed cache",
+		"size", cfg.Embedd.CacheSize,
+		"default_model", cfg.Embedd.DefaultModel,
+		"enabled", cfg.Embedd.CacheSize > 0)
+	h := &handlers{provider: cached, cache: cached}
 
 	if err := shared.Run("embedd", cfg.Embedd.Bind, h.register); err != nil {
 		slog.Error("server", "err", err)
@@ -53,10 +64,36 @@ func main() {
 
 type handlers struct {
 	provider embed.Provider
+	// cache is the same instance as provider when caching is enabled,
+	// kept as a typed pointer so /v1/embed/stats can expose hit-rate
+	// without type-asserting through the Provider interface. nil when
+	// CacheSize=0 (pass-through mode).
+	cache *embed.CachedProvider
 }
 
 func (h *handlers) register(r chi.Router) {
 	r.Post("/embed", h.handleEmbed)
+	r.Get("/embed/stats", h.handleStats)
+}
+
+// handleStats reports cache hits/misses + hit rate + size. Operators
+// use this to confirm the cache is doing its job (high hit rate) or
+// is sized wrong (low hit rate + many misses on a workload that
+// should have repeats).
+func (h *handlers) handleStats(w http.ResponseWriter, _ *http.Request) {
+	w.Header().Set("Content-Type", "application/json")
+	if h.cache == nil {
+		_ = json.NewEncoder(w).Encode(map[string]any{"enabled": false})
+		return
+	}
+	hits, misses := h.cache.Stats()
+	_ = json.NewEncoder(w).Encode(map[string]any{
+		"enabled":   true,
+		"hits":      hits,
+		"misses":    misses,
+		"hit_rate":  h.cache.HitRate(),
+		"size":      h.cache.Len(),
+	})
 }
 
 // embedRequest is the POST /embed body. Texts is the list to

cmd/embedd/main_test.go

@@ -143,6 +143,71 @@ func TestHandleEmbed_HappyPath_ProviderEcho(t *testing.T) {
 	}
 }
 
+func TestStatsEndpoint_CacheDisabled(t *testing.T) {
+	r := mountWithProvider(&stubProvider{})
+	srv := httptest.NewServer(r)
+	defer srv.Close()
+
+	resp, err := http.Get(srv.URL + "/embed/stats")
+	if err != nil {
+		t.Fatalf("GET: %v", err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		t.Errorf("status = %d, want 200", resp.StatusCode)
+	}
+	if !strings.Contains(getBody(t, resp), `"enabled":false`) {
+		t.Errorf("expected enabled:false when no cache wired")
+	}
+}
+
+func TestStatsEndpoint_CacheEnabled_TracksHitsAndMisses(t *testing.T) {
+	stub := &stubProvider{result: embed.Result{
+		Model:     "test-model",
+		Dimension: 3,
+		Vectors:   [][]float32{{0.1, 0.2, 0.3}},
+	}}
+	cache, err := embed.NewCachedProvider(stub, "test-model", 100)
+	if err != nil {
+		t.Fatalf("NewCachedProvider: %v", err)
+	}
+	h := &handlers{provider: cache, cache: cache}
+	r := chi.NewRouter()
+	h.register(r)
+	srv := httptest.NewServer(r)
+	defer srv.Close()
+
+	// First call → miss.
+	http.Post(srv.URL+"/embed", "application/json",
+		strings.NewReader(`{"texts":["hello"],"model":"test-model"}`))
+	// Second call same text → hit.
+	http.Post(srv.URL+"/embed", "application/json",
+		strings.NewReader(`{"texts":["hello"],"model":"test-model"}`))
+
+	resp, err := http.Get(srv.URL + "/embed/stats")
+	if err != nil {
+		t.Fatalf("stats GET: %v", err)
+	}
+	defer resp.Body.Close()
+	body := getBody(t, resp)
+	if !strings.Contains(body, `"enabled":true`) {
+		t.Errorf("expected enabled:true, body=%s", body)
+	}
+	if !strings.Contains(body, `"hits":1`) {
+		t.Errorf("expected hits:1, body=%s", body)
+	}
+	if !strings.Contains(body, `"misses":1`) {
+		t.Errorf("expected misses:1, body=%s", body)
+	}
+}
+
+func getBody(t *testing.T, resp *http.Response) string {
+	t.Helper()
+	buf := make([]byte, 4096)
+	n, _ := resp.Body.Read(buf)
+	return string(buf[:n])
+}
+
 func TestItoa(t *testing.T) {
 	cases := []struct {
 		in  int

go.mod

@@ -45,6 +45,7 @@ require (
 	github.com/goccy/go-json v0.10.6 // indirect
 	github.com/google/flatbuffers v25.12.19+incompatible // indirect
 	github.com/google/renameio v1.0.1 // indirect
+	github.com/hashicorp/golang-lru/v2 v2.0.7 // indirect
 	github.com/klauspost/compress v1.18.5 // indirect
 	github.com/klauspost/cpuid/v2 v2.3.0 // indirect
 	github.com/pierrec/lz4/v4 v4.1.26 // indirect

go.sum

@@ -84,6 +84,8 @@ github.com/google/renameio v1.0.1 h1:Lh/jXZmvZxb0BBeSY5VKEfidcbcbenKjZFzM/q0fSeU
 github.com/google/renameio v1.0.1/go.mod h1:t/HQoYBZSsWSNK35C6CO/TpPLDVWvxOHboWUAweKUpk=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
 github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
+github.com/hashicorp/golang-lru/v2 v2.0.7 h1:a+bsQ5rvGLjzHuww6tVxozPZFVghXaHOwFs4luLUK2k=
+github.com/hashicorp/golang-lru/v2 v2.0.7/go.mod h1:QeFd9opnmA6QUJc5vARoKUSoFhyfM2/ZepoAG6RGpeM=
 github.com/klauspost/compress v1.18.5 h1:/h1gH5Ce+VWNLSWqPzOVn6XBO+vJbCNGvjoaGBFW2IE=
 github.com/klauspost/compress v1.18.5/go.mod h1:cwPg85FWrGar70rWktvGQj8/hthj3wpl0PGDogxkrSQ=
 github.com/klauspost/cpuid/v2 v2.3.0 h1:S4CRMLnYUhGeDFDqkGriYKdfoFlDnMtqTiI/sFzhA9Y=

internal/embed/cached.go

@@ -0,0 +1,180 @@
+// cached.go — LRU caching wrapper around a Provider.
+//
+// Memoizes (effective_model, sha256(text)) → []float32. Repeat
+// queries return the stored vector without round-tripping to the
+// upstream embedding service. Per-text caching: a batch with
+// mixed hit/miss only fetches the misses, then merges the result
+// preserving the caller's input order.
+//
+// Memory budget: ~3 KB per entry at d=768. 10K-entry default ≈ 30 MB
+// — small enough for any realistic embedd deployment. Operators
+// raising the cap should weigh memory headroom against expected
+// hit rate.
+//
+// Why this exists: 500K staffing test (memory project_golang_lakehouse)
+// showed that the staffing co-pilot replays many of the same query
+// texts ("forklift driver", "welder Chicago", etc.). Caching them
+// drops repeat-query cost from ~50ms (Ollama round-trip) to <1µs
+// (LRU hit). Real production win documented in feedback_meta_index_vision.
+
+package embed
+
+import (
+	"context"
+	"crypto/sha256"
+	"encoding/hex"
+	"fmt"
+	"sync/atomic"
+
+	lru "github.com/hashicorp/golang-lru/v2"
+)
+
+// CachedProvider wraps a Provider with an LRU cache keyed on
+// (effective_model, sha256(text)). Thread-safe.
+type CachedProvider struct {
+	inner        Provider
+	defaultModel string
+	cache        *lru.Cache[string, []float32]
+	hits         atomic.Int64
+	misses       atomic.Int64
+}
+
+// NewCachedProvider wraps inner with an LRU cache of the given size.
+//
+// defaultModel is used to resolve cache keys when a request leaves
+// the model field empty: a request for model="" is treated as
+// model=defaultModel for cache-key purposes, so callers that mix
+// "" and the explicit model name still hit the same cache entry.
+//
+// Caller-side panic protection: size <= 0 is treated as "no cache"
+// — every Embed call passes through. Avoids forcing operators to
+// understand LRU sizing to disable.
+func NewCachedProvider(inner Provider, defaultModel string, size int) (*CachedProvider, error) {
+	if size <= 0 {
+		// Sentinel: nil cache means pass-through. NewCachedProvider
+		// stays callable so the wiring layer can always wrap.
+		return &CachedProvider{inner: inner, defaultModel: defaultModel}, nil
+	}
+	cache, err := lru.New[string, []float32](size)
+	if err != nil {
+		return nil, fmt.Errorf("embed cache init: %w", err)
+	}
+	return &CachedProvider{
+		inner:        inner,
+		defaultModel: defaultModel,
+		cache:        cache,
+	}, nil
+}
+
+// Embed returns vectors for texts, memoizing per (model, text). On
+// a batch with mixed hits/misses, only the misses round-trip to
+// inner; the result preserves caller ordering.
+func (c *CachedProvider) Embed(ctx context.Context, texts []string, model string) (Result, error) {
+	// Pass-through when caching disabled.
+	if c.cache == nil {
+		return c.inner.Embed(ctx, texts, model)
+	}
+
+	if len(texts) == 0 {
+		return Result{}, ErrEmptyTexts
+	}
+
+	effectiveModel := model
+	if effectiveModel == "" {
+		effectiveModel = c.defaultModel
+	}
+
+	// Pass 1: cache lookup; collect misses preserving original index
+	// so we can write the upstream result back into the right slots.
+	out := make([][]float32, len(texts))
+	missTexts := make([]string, 0, len(texts))
+	missIdx := make([]int, 0, len(texts))
+	for i, t := range texts {
+		key := cacheKey(effectiveModel, t)
+		if v, ok := c.cache.Get(key); ok {
+			out[i] = v
+			c.hits.Add(1)
+			continue
+		}
+		missTexts = append(missTexts, t)
+		missIdx = append(missIdx, i)
+		c.misses.Add(1)
+	}
+
+	// All hits — synthesize the result without an upstream call.
+	// Use the effective model + the first cached vector's length
+	// for the response. Every cached vector for the same model has
+	// the same dimension by construction (Provider guarantees it).
+	if len(missTexts) == 0 {
+		dim := 0
+		if len(out) > 0 && len(out[0]) > 0 {
+			dim = len(out[0])
+		}
+		return Result{
+			Model:     effectiveModel,
+			Dimension: dim,
+			Vectors:   out,
+		}, nil
+	}
+
+	// Pass 2: fetch the misses, populate cache, merge.
+	res, err := c.inner.Embed(ctx, missTexts, model)
+	if err != nil {
+		return Result{}, err
+	}
+	if len(res.Vectors) != len(missTexts) {
+		return Result{}, fmt.Errorf("embed cache: provider returned %d vectors for %d miss texts",
+			len(res.Vectors), len(missTexts))
+	}
+	// Cache under effectiveModel (request-derived), NOT res.Model
+	// (upstream-derived). Future requests with the same input shape
+	// — same explicit model OR the same "" — must hit the same key.
+	// Only ours is predictable from the request; upstream's resolution
+	// can drift if the upstream's default changes.
+	for j, t := range missTexts {
+		out[missIdx[j]] = res.Vectors[j]
+		c.cache.Add(cacheKey(effectiveModel, t), res.Vectors[j])
+	}
+
+	return Result{
+		Model:     res.Model,
+		Dimension: res.Dimension,
+		Vectors:   out,
+	}, nil
+}
+
+// Stats returns lifetime hit + miss counts for this cache. Atomic
+// reads — no locking. Useful for /health, /metrics, or operator
+// dashboards.
+func (c *CachedProvider) Stats() (hits, misses int64) {
+	return c.hits.Load(), c.misses.Load()
+}
+
+// HitRate returns the fraction of requests served from cache.
+// Returns 0.0 when no requests have been served (avoids NaN).
+func (c *CachedProvider) HitRate() float64 {
+	h := c.hits.Load()
+	m := c.misses.Load()
+	total := h + m
+	if total == 0 {
+		return 0.0
+	}
+	return float64(h) / float64(total)
+}
+
+// Len returns the current entry count in the cache. Returns 0 when
+// caching is disabled.
+func (c *CachedProvider) Len() int {
+	if c.cache == nil {
+		return 0
+	}
+	return c.cache.Len()
+}
+
+// cacheKey is "<model>:<sha256(text)>". sha256 collapses long texts
+// to a fixed-size key; model prefix scopes cache to one model so
+// callers using multiple models don't get cross-contamination.
+func cacheKey(model, text string) string {
+	h := sha256.Sum256([]byte(text))
+	return model + ":" + hex.EncodeToString(h[:])
+}

internal/embed/cached_test.go

@@ -0,0 +1,350 @@
+package embed
+
+import (
+	"context"
+	"errors"
+	"sync"
+	"testing"
+)
+
+// stubProvider records every call so tests can verify caching
+// behavior (call counts, request texts) without spinning up Ollama.
+type stubProvider struct {
+	mu        sync.Mutex
+	model     string
+	dim       int
+	callCount int
+	lastTexts []string
+	err       error
+	// vectorFn returns a deterministic vector for a given text so
+	// cache hits assert byte-equality, not just shape.
+	vectorFn func(text string) []float32
+}
+
+func (s *stubProvider) Embed(_ context.Context, texts []string, model string) (Result, error) {
+	s.mu.Lock()
+	s.callCount++
+	s.lastTexts = append([]string(nil), texts...)
+	s.mu.Unlock()
+	if s.err != nil {
+		return Result{}, s.err
+	}
+	out := make([][]float32, len(texts))
+	for i, t := range texts {
+		out[i] = s.vectorFn(t)
+	}
+	resolvedModel := model
+	if resolvedModel == "" {
+		resolvedModel = s.model
+	}
+	return Result{Model: resolvedModel, Dimension: s.dim, Vectors: out}, nil
+}
+
+func deterministicVector(t string) []float32 {
+	v := make([]float32, 4)
+	for i := range v {
+		v[i] = float32(int(t[0]) + i)
+	}
+	return v
+}
+
+func newStub() *stubProvider {
+	return &stubProvider{
+		model:    "test-model",
+		dim:      4,
+		vectorFn: deterministicVector,
+	}
+}
+
+func TestCachedProvider_PassThroughWhenSizeZero(t *testing.T) {
+	stub := newStub()
+	c, err := NewCachedProvider(stub, "test-model", 0)
+	if err != nil {
+		t.Fatalf("NewCachedProvider: %v", err)
+	}
+	for i := 0; i < 3; i++ {
+		_, err := c.Embed(context.Background(), []string{"hello"}, "")
+		if err != nil {
+			t.Fatalf("Embed: %v", err)
+		}
+	}
+	if stub.callCount != 3 {
+		t.Errorf("size=0 should pass through every call, got callCount=%d", stub.callCount)
+	}
+	if c.Len() != 0 {
+		t.Errorf("Len()=%d, want 0 when caching disabled", c.Len())
+	}
+}
+
+func TestCachedProvider_HitOnRepeatText(t *testing.T) {
+	stub := newStub()
+	c, err := NewCachedProvider(stub, "test-model", 100)
+	if err != nil {
+		t.Fatalf("NewCachedProvider: %v", err)
+	}
+
+	// First call — miss, hits upstream.
+	r1, err := c.Embed(context.Background(), []string{"hello"}, "test-model")
+	if err != nil {
+		t.Fatalf("first: %v", err)
+	}
+	if stub.callCount != 1 {
+		t.Errorf("first call should hit upstream, callCount=%d", stub.callCount)
+	}
+
+	// Second call — hit, no upstream call.
+	r2, err := c.Embed(context.Background(), []string{"hello"}, "test-model")
+	if err != nil {
+		t.Fatalf("second: %v", err)
+	}
+	if stub.callCount != 1 {
+		t.Errorf("second call should hit cache, callCount=%d (want unchanged 1)", stub.callCount)
+	}
+
+	// Result should be byte-equal.
+	if len(r1.Vectors[0]) != len(r2.Vectors[0]) {
+		t.Errorf("vector dim differs: %d vs %d", len(r1.Vectors[0]), len(r2.Vectors[0]))
+	}
+	for i := range r1.Vectors[0] {
+		if r1.Vectors[0][i] != r2.Vectors[0][i] {
+			t.Errorf("vector[%d]: %v vs %v", i, r1.Vectors[0][i], r2.Vectors[0][i])
+		}
+	}
+
+	hits, misses := c.Stats()
+	if hits != 1 || misses != 1 {
+		t.Errorf("hits/misses = %d/%d, want 1/1", hits, misses)
+	}
+}
+
+func TestCachedProvider_MixedBatchOnlyFetchesMisses(t *testing.T) {
+	stub := newStub()
+	c, _ := NewCachedProvider(stub, "test-model", 100)
+
+	// Prime cache with one text.
+	c.Embed(context.Background(), []string{"alpha"}, "test-model")
+	if stub.callCount != 1 {
+		t.Fatalf("priming should produce 1 upstream call, got %d", stub.callCount)
+	}
+
+	// Mixed batch: alpha cached, beta + gamma fresh.
+	r, err := c.Embed(context.Background(),
+		[]string{"alpha", "beta", "gamma"}, "test-model")
+	if err != nil {
+		t.Fatalf("Embed: %v", err)
+	}
+
+	// Provider should have been called only with the misses.
+	if len(stub.lastTexts) != 2 {
+		t.Errorf("upstream called with %d texts, want 2 (only misses)", len(stub.lastTexts))
+	}
+	for _, lt := range stub.lastTexts {
+		if lt == "alpha" {
+			t.Errorf("upstream got 'alpha' but it was cached")
+		}
+	}
+
+	// Result preserves order: alpha first (cached), beta + gamma in
+	// their original positions.
+	if len(r.Vectors) != 3 {
+		t.Fatalf("got %d vectors, want 3", len(r.Vectors))
+	}
+	for i, txt := range []string{"alpha", "beta", "gamma"} {
+		want := deterministicVector(txt)
+		for j := range want {
+			if r.Vectors[i][j] != want[j] {
+				t.Errorf("vec[%d (%s)][%d] = %v, want %v",
+					i, txt, j, r.Vectors[i][j], want[j])
+			}
+		}
+	}
+}
+
+func TestCachedProvider_ModelKeyIsolation(t *testing.T) {
+	stub := newStub()
+	c, _ := NewCachedProvider(stub, "default-model", 100)
+
+	// Same text, different models → distinct cache entries.
+	c.Embed(context.Background(), []string{"shared"}, "model-A")
+	c.Embed(context.Background(), []string{"shared"}, "model-B")
+	if stub.callCount != 2 {
+		t.Errorf("different models should NOT share cache, got callCount=%d", stub.callCount)
+	}
+
+	// Repeat with model-A → hits cache.
+	c.Embed(context.Background(), []string{"shared"}, "model-A")
+	if stub.callCount != 2 {
+		t.Errorf("repeat of model-A should hit cache, got callCount=%d", stub.callCount)
+	}
+}
+
+func TestCachedProvider_EmptyModelResolvesToDefault(t *testing.T) {
+	stub := newStub()
+	c, _ := NewCachedProvider(stub, "the-default", 100)
+
+	// First call with model="" — caches under "the-default".
+	c.Embed(context.Background(), []string{"text"}, "")
+	if stub.callCount != 1 {
+		t.Fatalf("first call should hit upstream, got %d", stub.callCount)
+	}
+
+	// Second call with model="the-default" — should hit cache because
+	// the first call resolved "" → "the-default" for the key.
+	c.Embed(context.Background(), []string{"text"}, "the-default")
+	if stub.callCount != 1 {
+		t.Errorf("explicit default-model should hit cache populated by '' request, got callCount=%d", stub.callCount)
+	}
+}
+
+func TestCachedProvider_LRUEviction(t *testing.T) {
+	stub := newStub()
+	c, _ := NewCachedProvider(stub, "test-model", 2) // cap=2
+
+	// Fill: a, b
+	c.Embed(context.Background(), []string{"a"}, "")
+	c.Embed(context.Background(), []string{"b"}, "")
+	// c arrives — evicts a (LRU).
+	c.Embed(context.Background(), []string{"c"}, "")
+	if c.Len() > 2 {
+		t.Errorf("Len()=%d, want ≤2 with cap=2", c.Len())
+	}
+	if stub.callCount != 3 {
+		t.Errorf("callCount=%d, want 3 after a/b/c distinct", stub.callCount)
+	}
+
+	// Re-request a → miss (was evicted).
+	c.Embed(context.Background(), []string{"a"}, "")
+	if stub.callCount != 4 {
+		t.Errorf("evicted a should miss, callCount=%d want 4", stub.callCount)
+	}
+}
+
+func TestCachedProvider_PropagatesUpstreamError(t *testing.T) {
+	stub := &stubProvider{
+		model:    "test-model",
+		dim:      4,
+		vectorFn: deterministicVector,
+		err:      errors.New("upstream broken"),
+	}
+	c, _ := NewCachedProvider(stub, "test-model", 10)
+
+	_, err := c.Embed(context.Background(), []string{"hello"}, "")
+	if err == nil {
+		t.Fatal("expected upstream error to propagate")
+	}
+	// Cache should NOT have stored anything on error.
+	if c.Len() != 0 {
+		t.Errorf("error path should not populate cache, Len()=%d", c.Len())
+	}
+}
+
+func TestCachedProvider_AllHitsSynthesized(t *testing.T) {
+	stub := newStub()
+	c, _ := NewCachedProvider(stub, "test-model", 100)
+
+	// Prime
+	c.Embed(context.Background(), []string{"x", "y"}, "")
+	stub.mu.Lock()
+	stub.callCount = 0 // reset for clarity
+	stub.mu.Unlock()
+
+	// All-hits batch → no upstream call.
+	r, err := c.Embed(context.Background(), []string{"x", "y"}, "")
+	if err != nil {
+		t.Fatalf("Embed: %v", err)
+	}
+	if stub.callCount != 0 {
+		t.Errorf("all-hits batch should not hit upstream, callCount=%d", stub.callCount)
+	}
+	if r.Dimension != 4 {
+		t.Errorf("all-hits Dimension=%d, want 4 (from cached vector len)", r.Dimension)
+	}
+	if r.Model != "test-model" {
+		t.Errorf("all-hits Model=%q, want resolved 'test-model'", r.Model)
+	}
+}
+
+func TestCachedProvider_HitRate(t *testing.T) {
+	stub := newStub()
+	c, _ := NewCachedProvider(stub, "test-model", 100)
+
+	if c.HitRate() != 0.0 {
+		t.Errorf("initial HitRate=%v, want 0.0", c.HitRate())
+	}
+
+	c.Embed(context.Background(), []string{"a"}, "")    // miss
+	c.Embed(context.Background(), []string{"a"}, "")    // hit
+	c.Embed(context.Background(), []string{"a"}, "")    // hit
+
+	got := c.HitRate()
+	want := 2.0 / 3.0
+	if got < want-0.01 || got > want+0.01 {
+		t.Errorf("HitRate=%v, want %v", got, want)
+	}
+}
+
+func TestCachedProvider_EmptyTextsRejected(t *testing.T) {
+	stub := newStub()
+	c, _ := NewCachedProvider(stub, "test-model", 100)
+
+	_, err := c.Embed(context.Background(), []string{}, "")
+	if !errors.Is(err, ErrEmptyTexts) {
+		t.Errorf("empty texts should return ErrEmptyTexts, got %v", err)
+	}
+}
+
+func TestCachedProvider_ConcurrentSafe(t *testing.T) {
+	stub := newStub()
+	c, _ := NewCachedProvider(stub, "test-model", 100)
+
+	// 50 goroutines × 100 calls each, mostly the same text.
+	var wg sync.WaitGroup
+	for g := 0; g < 50; g++ {
+		wg.Add(1)
+		go func(id int) {
+			defer wg.Done()
+			for i := 0; i < 100; i++ {
+				_, err := c.Embed(context.Background(), []string{"shared"}, "")
+				if err != nil {
+					t.Errorf("concurrent embed err: %v", err)
+					return
+				}
+			}
+		}(g)
+	}
+	wg.Wait()
+
+	hits, misses := c.Stats()
+	if hits+misses != 5000 {
+		t.Errorf("total calls = %d, want 5000", hits+misses)
+	}
+	// First call is a guaranteed miss; subsequent should mostly be
+	// hits. With 50 concurrent goroutines hitting an empty cache,
+	// up to ~50 first-arrivers may miss before the first one writes.
+	if misses > 100 {
+		t.Errorf("misses=%d unreasonably high; cache concurrency broken?", misses)
+	}
+}
+
+func TestCacheKey_StableAndDistinct(t *testing.T) {
+	a := cacheKey("m1", "text-a")
+	a2 := cacheKey("m1", "text-a")
+	b := cacheKey("m1", "text-b")
+	c := cacheKey("m2", "text-a")
+
+	if a != a2 {
+		t.Errorf("cache key not stable: %q != %q", a, a2)
+	}
+	if a == b {
+		t.Errorf("different texts produced same key: %q", a)
+	}
+	if a == c {
+		t.Errorf("different models produced same key: %q", a)
+	}
+	// Key shape: model + ":" + 64-hex-chars sha256.
+	const sha256HexLen = 64
+	want := len("m1") + 1 + sha256HexLen
+	if len(a) != want {
+		t.Errorf("key len=%d, want %d", len(a), want)
+	}
+}

internal/shared/config.go

@@ -65,11 +65,14 @@ type GatewayConfig struct {
 // EmbeddConfig drives the embed service. ProviderURL points at the
 // embedding backend (Ollama in G2, possibly OpenAI/Voyage in G3+).
 // DefaultModel is what gets used when callers don't specify a
-// model in their request body.
+// model in their request body. CacheSize is the LRU cache cap on
+// (model, sha256(text)) → vector lookups; 0 disables caching.
+// Default 10000 entries ≈ 30 MiB at d=768.
 type EmbeddConfig struct {
 	Bind         string `toml:"bind"`
 	ProviderURL  string `toml:"provider_url"`
 	DefaultModel string `toml:"default_model"`
+	CacheSize    int    `toml:"cache_size"`
 }
 
 // VectordConfig adds vectord-specific knobs. StoragedURL is
@@ -153,6 +156,7 @@ func DefaultConfig() Config {
 			Bind:         "127.0.0.1:3216",
 			ProviderURL:  "http://localhost:11434", // local Ollama
 			DefaultModel: "nomic-embed-text",
+			CacheSize:    10_000, // ~30 MiB at d=768; set to 0 to disable
 		},
 		Queryd: QuerydConfig{
 			Bind:         "127.0.0.1:3214",