Feat v0.8.3 vector column (#70)

Co-authored-by: Jeffrey Smith <jasafpro@gmail.com>
Co-committed-by: Jeffrey Smith <jasafpro@gmail.com>

CHANGELOG.md

@@ -2,6 +2,40 @@
 
 All notable changes to Armature are documented here.
 
+## v0.8.3 — Vector Column Type
+
+Extensions can now declare vector columns and perform similarity search.
+Three-tier progressive enhancement: native pgvector on Postgres, JSONB
+fallback without pgvector, TEXT fallback on SQLite.
+
+**Manifest: `db_tables` vector columns**
+
+- Declare `"vector(N)"` as a column type (N = dimension, 1–4096).
+- On Postgres with pgvector: native `vector(N)` type with auto-created
+  HNSW index (`vector_cosine_ops`).
+- On Postgres without pgvector: `JSONB` column.
+- On SQLite: `TEXT` column (JSON-encoded float arrays).
+
+**Starlark API**
+
+- `db.query_similar(table, column, vector=[], limit=10, filters={}, metric="cosine")`
+  — returns rows ordered by ascending cosine distance with injected `_distance` key.
+- `db.insert()` now accepts list values (serialized as JSON strings) for
+  vector column storage.
+
+**Internal**
+
+- Modified: `handlers/ext_db_schema.go` — `parseVectorDim`, `mapColType`
+  gains `hasPgvector` parameter, HNSW index creation for vector columns.
+- Modified: `sandbox/db_module.go` — `HasPgvector` in `DBModuleConfig`,
+  list support in `starlarkToGoValue`, `dbQuerySimilar` with pgvector and
+  fallback paths, `cosineDistance` helper.
+- Modified: `sandbox/runner.go` — wire `HasPgvector` from capabilities.
+- Modified: `handlers/extensions.go` — `SetCapabilities` on `ExtensionHandler`.
+- Modified: `server/main.go` — wire capabilities to extension handler.
+- Updated: `docs/STARLARK-REFERENCE.md` — vector similarity section.
+- New tests: 5 schema tests + 8 db module tests (13 total).
+
 ## v0.8.2 — Capability Negotiation
 
 Extensions declare environment requirements in their manifest. The kernel

ROADMAP.md

@@ -1,6 +1,6 @@
 # Armature — Roadmap
 
-## Current: v0.8.1 — Workspace Module
+## Current: v0.8.4 — Documentation Refresh + Surface Sizing Fix
 
 Self-hosted extensible platform kernel. Auth, identity, packages, Starlark
 sandbox, storage, realtime, and ops are kernel primitives. Everything else
@@ -424,15 +424,43 @@ New: `handlers/capabilities.go`, `handlers/capabilities_test.go`,
 `sandbox/settings_module_test.go`. Modified: install handler, bundled
 handler, settings module, runner, main.go. 18 new tests.
 
-**v0.8.3 — Vector Column Type**
+**v0.8.3 — Vector Column Type** ✅
 
 `"vector(N)"` in `db_tables` manifest. Maps to native `vector(N)` on
 PG+pgvector, `JSONB` on PG without, `TEXT` on SQLite. New `db.query_similar()`
-with dual-path dispatch (native operator vs Go-side brute force).
+with dual-path dispatch (native operator vs Go-side brute force). Auto-created
+HNSW index on pgvector. `starlarkToGoValue` extended with list→JSON.
 
-Modified: `ext_db_schema.go`, `db_module.go`.
+New: `docs/DESIGN-vector-column.md`. Modified: `ext_db_schema.go`,
+`db_module.go`, `runner.go`, `extensions.go`, `main.go`. 13 new tests.
 
-**v0.8.4 — Extension Composability**
+**v0.8.4 — Documentation Refresh + Surface Sizing Fix**
+
+Two unrelated items bundled for a clean changeset before composability.
+
+*Documentation refresh:* Starlark/extension docs have drifted since v0.7.5.
+Modules added or significantly changed but not fully documented:
+- `workspace` module (v0.8.1) — undocumented in STARLARK-REFERENCE.md
+- `permissions` module (v0.7.7) — undocumented in STARLARK-REFERENCE.md
+- `settings.has_capability()` (v0.8.2) — not in settings section
+- `db.query_similar()` vector section (v0.8.3) — added but needs review pass
+- EXTENSION-GUIDE.md — needs `capabilities` manifest block, `vector(N)` column
+  type, `user_permissions`, `gate_permission`
+- TUTORIAL-FIRST-EXTENSION.md — review for v0.7+ accuracy
+
+*Surface sizing bug:* Extension surfaces (`height: 100%` + `overflow: hidden`)
+don't account for the 44px shell topbar. Content is cut off at the bottom
+with no scroll. Affected: chat, notes, dashboard, schedules, and likely all
+extension surfaces. Root cause: `.extension-mount` is `flex: 1` within the
+viewport column, but extensions set `height: 100%` on their root element
+which inherits the mount's full height — the topbar's 44px is already
+subtracted by flex, but the extension's internal layout overflows because
+child containers also use `height: 100%` + `overflow: hidden`, clipping
+the last 44px of content. Fix: audit all extension root CSS and ensure
+inner scroll containers use `flex: 1; min-height: 0; overflow: auto`
+instead of `height: 100%; overflow: hidden`.
+
+**v0.8.5 — Extension Composability**
 
 Design doc: `docs/DESIGN-extension-composability.md`
 
@@ -616,6 +644,7 @@ These are candidates, not commitments. Each requires a design doc.
 | `workspace` for real filesystem | Flat blob store can't serve git/compilers/ffmpeg. Managed disk paths are the explicit escape hatch. |
 | Capability negotiation at install | Fail loud with actionable message, not silently at runtime. |
 | Vector column with three-tier fallback | Works everywhere, works fast with pgvector. Documentation > hard gates. |
+| Docs refresh before composability | 8 versions of module additions (v0.7.5–v0.8.3) without a docs pass. Ship accurate docs before adding more API surface. |
 | Sidecar deferred to v0.10.x | HTTP module covers external APIs. Local compute is a v0.10.x concern. Sidecars connect inward (like cluster nodes), not spawned outward (no k8s RBAC needed). |
 | No second scripting runtime | Starlark for glue, HTTP for external compute, sidecars for local compute. Three tiers with clear boundaries. Lua/WASM middle tier rejected — Turing-complete enough to be dangerous, not isolated enough to be trustworthy. |
 | Reference extensions prove the platform | First-party `.pkg` files, not kernel code. Uninstallable. Dogfooding. |

VERSION

@@ -1 +1 @@
-0.8.2
+0.8.3

docs/DESIGN-vector-column.md

@@ -0,0 +1,97 @@
+# Design: Vector Column Type (v0.8.3)
+
+## Problem
+
+Extensions building semantic search, RAG, or recommendation features need to
+store and query high-dimensional vectors (embeddings). Without kernel-level
+support, each extension would need to reinvent storage, serialization, and
+similarity search — duplicating effort and missing the pgvector optimization
+path.
+
+## Design
+
+### Three-tier progressive enhancement
+
+| Backend | Column DDL | Storage | Search |
+|---------|-----------|---------|--------|
+| Postgres + pgvector | `vector(N)` + HNSW index | Native vector type | `<=>` operator (index-backed) |
+| Postgres (no pgvector) | `JSONB` | JSON array | Go-side cosine distance |
+| SQLite | `TEXT` | JSON string | Go-side cosine distance |
+
+Extensions declare `"vector(N)"` in their manifest `db_tables` block.
+The kernel maps this to the appropriate SQL type at install time based on
+detected capabilities.
+
+### Manifest example
+
+```json
+{
+  "db_tables": {
+    "documents": {
+      "columns": {
+        "title": "text",
+        "embedding": "vector(384)"
+      }
+    }
+  },
+  "capabilities": {
+    "optional": ["pgvector"]
+  }
+}
+```
+
+### API surface
+
+```python
+# Insert (vector as list)
+db.insert("documents", {"title": "hello", "embedding": [0.1, 0.2, ...]})
+
+# Similarity search
+rows = db.query_similar(
+    "documents", "embedding",
+    vector=[0.1, 0.2, ...],
+    limit=10,
+    filters={"active": True},
+    metric="cosine"
+)
+# → [{..., "_distance": 0.023}, ...]
+```
+
+### Dispatch paths
+
+**pgvector path** — SQL-side computation with index:
+```sql
+SELECT *, (embedding <=> $1::vector) AS _distance
+FROM ext_pkg_documents
+WHERE ... ORDER BY embedding <=> $1::vector LIMIT $2
+```
+
+**Fallback path** — Go-side computation:
+1. `SELECT * FROM table WHERE filters LIMIT 1000`
+2. Parse each row's vector column from JSON
+3. Compute cosine distance in Go
+4. Sort by distance, return top N with `_distance` injected
+
+### Dimension validation
+
+- Manifest: 1 ≤ N ≤ 4096 (validated by `parseVectorDim`)
+- Insert-time: no dimension validation (store whatever list is given)
+- Query-time: dimension mismatches produce distance = 1.0 (treated as unrelated)
+
+### Performance characteristics
+
+| Path | 1K rows | 10K rows | 100K rows |
+|------|---------|----------|-----------|
+| pgvector (HNSW) | <1ms | <5ms | <10ms |
+| Fallback (Go) | <10ms | ~100ms | Not recommended |
+
+Fallback caps at 1000 rows fetched. Extensions needing large-scale similarity
+search should declare `capabilities.optional: ["pgvector"]` and degrade
+gracefully.
+
+## Limitations
+
+- Only cosine distance metric (v0.8.3). L2 / inner product can be added later.
+- No dimension validation at insert time.
+- Fallback path fetches at most 1000 rows — not suitable for large datasets.
+- HNSW index is only created for pgvector backends.

docs/STARLARK-REFERENCE.md

@@ -136,6 +136,33 @@ results = db.query_batch([
 # Each query spec supports: table (required), filters, order, limit, before, after, search_like
 ```
 
+#### Vector similarity search
+
+```python
+# Find rows with the most similar embeddings (cosine distance)
+rows = db.query_similar(
+    "documents",        # table name
+    "embedding",        # vector column name
+    vector=[0.1, 0.2, ...],  # query vector (list of floats)
+    limit=10,           # max results (default 10, max 100)
+    filters={"active": True},  # optional equality filters
+    metric="cosine",    # only "cosine" supported
+)
+# Returns rows ordered by ascending _distance (0.0 = identical, 1.0 = orthogonal)
+# Each row dict includes an injected "_distance" float key.
+```
+
+Vector columns are declared as `"vector(N)"` in the manifest `db_tables` block
+(N = dimension, 1–4096). Storage varies by backend:
+
+| Backend | Column type | Search |
+|---------|-------------|--------|
+| Postgres + pgvector | `vector(N)` with HNSW index | Native `<=>` operator |
+| Postgres (no pgvector) | `JSONB` | Go-side cosine computation |
+| SQLite | `TEXT` | Go-side cosine computation |
+
+Insert vectors as lists: `db.insert("docs", {"embedding": [0.1, 0.2, 0.3]})`.
+
 #### Write operations
 
 ```python

server/handlers/ext_db_schema.go

@@ -13,20 +13,38 @@ import (
 	"fmt"
 	"log"
 	"regexp"
+	"strconv"
 	"strings"
 
 	"armature/store"
 )
 
+// vectorDimRegex matches manifest type strings like "vector(384)".
+var vectorDimRegex = regexp.MustCompile(`^vector\((\d+)\)$`)
+
 // validSchemaIdentifier matches safe SQL identifiers for table and column names.
 var validSchemaIdentifier = regexp.MustCompile(`^[a-z][a-z0-9_]{0,62}$`)
 
 // TableDef describes a single extension-owned table as declared in a manifest.
 type TableDef struct {
-	Columns map[string]string // colName → manifest type ("text","int","real","bool","timestamp")
+	Columns map[string]string // colName → manifest type ("text","int","real","bool","timestamp","vector(N)")
 	Indexes [][]string        // each inner slice is an ordered list of column names for one index
 }
 
+// parseVectorDim extracts the dimension N from a "vector(N)" type string.
+// Returns (N, true) for valid dimensions 1..4096, or (0, false) otherwise.
+func parseVectorDim(typStr string) (int, bool) {
+	m := vectorDimRegex.FindStringSubmatch(strings.ToLower(typStr))
+	if m == nil {
+		return 0, false
+	}
+	n, err := strconv.Atoi(m[1])
+	if err != nil || n < 1 || n > 4096 {
+		return 0, false
+	}
+	return n, true
+}
+
 // ParseDBTables extracts the "db_tables" key from a package manifest map.
 // Returns the map of logicalName→TableDef and ok=true when the key is present
 // and produces at least one valid table definition.
@@ -93,8 +111,22 @@ func ParseDBTables(manifest map[string]any) (map[string]TableDef, bool) {
 }
 
 // mapColType maps a manifest type string to a SQL column type for the given dialect.
-func mapColType(typStr string, isPostgres bool) string {
+// hasPgvector indicates whether the pgvector extension is available on Postgres.
-	switch strings.ToLower(typStr) {
+func mapColType(typStr string, isPostgres bool, hasPgvector bool) string {
+	lower := strings.ToLower(typStr)
+
+	// Check for vector type first.
+	if _, ok := parseVectorDim(lower); ok {
+		if isPostgres && hasPgvector {
+			return lower // native pgvector type, e.g. "vector(384)"
+		}
+		if isPostgres {
+			return "JSONB" // structured fallback without pgvector
+		}
+		return "TEXT" // SQLite fallback — JSON string
+	}
+
+	switch lower {
 	case "int", "integer":
 		return "INTEGER"
 	case "real", "float":
@@ -133,6 +165,8 @@ func createdAtColDef(isPostgres bool) string {
 // in the provided map.  Each successfully created table is registered in the
 // ext_data_tables catalog via stores.ExtData.
 //
+// hasPgvector enables native pgvector column types and HNSW indexes when true.
+//
 // Errors from individual table creation are returned immediately (fail-fast).
 // Index creation failures are logged but non-fatal.
 // Catalog registration failures are logged but non-fatal.
@@ -143,6 +177,7 @@ func CreateExtTables(
 	stores store.Stores,
 	packageID string,
 	tables map[string]TableDef,
+	hasPgvector bool,
 ) error {
 	if db == nil {
 		return nil
@@ -153,7 +188,7 @@ func CreateExtTables(
 		// Build column list: id first, user columns, created_at last.
 		cols := []string{"id TEXT PRIMARY KEY"}
 		for col, typ := range td.Columns {
-			cols = append(cols, col+" "+mapColType(typ, isPostgres))
+			cols = append(cols, col+" "+mapColType(typ, isPostgres, hasPgvector))
 		}
 		cols = append(cols, createdAtColDef(isPostgres))
 
@@ -174,6 +209,21 @@ func CreateExtTables(
 			}
 		}
 
+		// Create HNSW indexes for vector columns when pgvector is available.
+		if isPostgres && hasPgvector {
+			for col, typ := range td.Columns {
+				if _, ok := parseVectorDim(strings.ToLower(typ)); ok {
+					hnswName := fmt.Sprintf("idx_%s_%s_hnsw", physical, col)
+					hnswDDL := fmt.Sprintf(
+						"CREATE INDEX IF NOT EXISTS %s ON %s USING hnsw (%s vector_cosine_ops)",
+						hnswName, physical, col)
+					if _, err := db.ExecContext(ctx, hnswDDL); err != nil {
+						log.Printf("[ext-db] hnsw index create failed (%s): %v", hnswName, err)
+					}
+				}
+			}
+		}
+
 		// Record logical name in catalog.
 		if stores.ExtData != nil {
 			if err := stores.ExtData.RegisterTable(ctx, packageID, logicalName); err != nil {
@@ -245,6 +295,7 @@ func MigrateExtTables(
 	stores store.Stores,
 	packageID string,
 	newTables map[string]TableDef,
+	hasPgvector bool,
 ) ([]string, error) {
 	if db == nil {
 		return nil, nil
@@ -270,7 +321,7 @@ func MigrateExtTables(
 		if !existingTableNames[logicalName] {
 			// New table — full creation
 			if err := CreateExtTables(ctx, db, isPostgres, stores, packageID,
-				map[string]TableDef{logicalName: td}); err != nil {
+				map[string]TableDef{logicalName: td}, hasPgvector); err != nil {
 				return changes, fmt.Errorf("create new table %s: %w", physical, err)
 			}
 			changes = append(changes, fmt.Sprintf("created table %s", physical))
@@ -287,7 +338,7 @@ func MigrateExtTables(
 			if existingCols[col] {
 				continue
 			}
-			sqlType := mapColType(typ, isPostgres)
+			sqlType := mapColType(typ, isPostgres, hasPgvector)
 			alterDDL := fmt.Sprintf("ALTER TABLE %s ADD COLUMN %s %s", physical, col, sqlType)
 			if _, err := db.ExecContext(ctx, alterDDL); err != nil {
 				return changes, fmt.Errorf("add column %s.%s: %w", physical, col, err)

server/handlers/ext_db_schema_test.go

@@ -145,7 +145,7 @@ func TestMapColType_SQLite(t *testing.T) {
 		{"unknown", "TEXT"},
 	}
 	for _, c := range cases {
-		got := mapColType(c.in, false)
+		got := mapColType(c.in, false, false)
 		if got != c.want {
 			t.Errorf("mapColType(%q, sqlite) = %q, want %q", c.in, got, c.want)
 		}
@@ -153,13 +153,13 @@ func TestMapColType_SQLite(t *testing.T) {
 }
 
 func TestMapColType_Postgres(t *testing.T) {
-	if mapColType("bool", true) != "BOOLEAN" {
+	if mapColType("bool", true, false) != "BOOLEAN" {
 		t.Error("expected BOOLEAN for bool in postgres")
 	}
-	if mapColType("timestamp", true) != "TIMESTAMPTZ" {
+	if mapColType("timestamp", true, false) != "TIMESTAMPTZ" {
 		t.Error("expected TIMESTAMPTZ for timestamp in postgres")
 	}
-	if mapColType("int", true) != "INTEGER" {
+	if mapColType("int", true, false) != "INTEGER" {
 		t.Error("expected INTEGER for int in postgres")
 	}
 }
@@ -189,7 +189,7 @@ func TestCreateExtTables_CreatesTable(t *testing.T) {
 			Columns: map[string]string{"message": "text", "count": "int"},
 		},
 	}
-	if err := CreateExtTables(ctx, db, false, storesWithExtData(m), "test-pkg", tables); err != nil {
+	if err := CreateExtTables(ctx, db, false, storesWithExtData(m), "test-pkg", tables, false); err != nil {
 		t.Fatalf("CreateExtTables: %v", err)
 	}
 
@@ -212,7 +212,7 @@ func TestCreateExtTables_WithIndexes(t *testing.T) {
 			Indexes: [][]string{{"user_id"}, {"user_id", "kind"}},
 		},
 	}
-	if err := CreateExtTables(ctx, db, false, storesWithExtData(m), "idx-pkg", tables); err != nil {
+	if err := CreateExtTables(ctx, db, false, storesWithExtData(m), "idx-pkg", tables, false); err != nil {
 		t.Fatalf("CreateExtTables: %v", err)
 	}
 
@@ -243,7 +243,7 @@ func TestCreateExtTables_CatalogRegistration(t *testing.T) {
 		"logs":   {Columns: map[string]string{"msg": "text"}},
 		"errors": {Columns: map[string]string{"detail": "text"}},
 	}
-	if err := CreateExtTables(ctx, db, false, storesWithExtData(m), "cat-pkg", tables); err != nil {
+	if err := CreateExtTables(ctx, db, false, storesWithExtData(m), "cat-pkg", tables, false); err != nil {
 		t.Fatalf("CreateExtTables: %v", err)
 	}
 
@@ -259,7 +259,7 @@ func TestCreateExtTables_NilDBIsNoop(t *testing.T) {
 	// Should not panic or error.
 	err := CreateExtTables(ctx, nil, false, storesWithExtData(m), "pkg", map[string]TableDef{
 		"t": {Columns: map[string]string{"x": "text"}},
-	})
+	}, false)
 	if err != nil {
 		t.Errorf("expected nil error for nil db, got: %v", err)
 	}
@@ -276,7 +276,7 @@ func TestDropExtTables_DropsAndClearsCatalog(t *testing.T) {
 	tables := map[string]TableDef{
 		"items": {Columns: map[string]string{"name": "text"}},
 	}
-	if err := CreateExtTables(ctx, db, false, storesWithExtData(m), "drop-pkg", tables); err != nil {
+	if err := CreateExtTables(ctx, db, false, storesWithExtData(m), "drop-pkg", tables, false); err != nil {
 		t.Fatalf("CreateExtTables: %v", err)
 	}
 
@@ -309,7 +309,7 @@ func TestDropExtTables_MultipleTables(t *testing.T) {
 		"alpha": {Columns: map[string]string{"v": "text"}},
 		"beta":  {Columns: map[string]string{"v": "text"}},
 	}
-	CreateExtTables(ctx, db, false, storesWithExtData(m), "multi-pkg", tables)
+	CreateExtTables(ctx, db, false, storesWithExtData(m), "multi-pkg", tables, false)
 	DropExtTables(ctx, db, storesWithExtData(m), "multi-pkg")
 
 	for _, name := range []string{"ext_multi_pkg_alpha", "ext_multi_pkg_beta"} {
@@ -331,7 +331,7 @@ func TestCreateExtTables_DDLContainsAutoColumns(t *testing.T) {
 	ctx := context.Background()
 
 	tables := map[string]TableDef{"empty": {Columns: map[string]string{}}}
-	if err := CreateExtTables(ctx, db, false, storesWithExtData(m), "auto-pkg", tables); err != nil {
+	if err := CreateExtTables(ctx, db, false, storesWithExtData(m), "auto-pkg", tables, false); err != nil {
 		t.Fatalf("CreateExtTables: %v", err)
 	}
 
@@ -348,3 +348,77 @@ func TestCreateExtTables_DDLContainsAutoColumns(t *testing.T) {
 		t.Error("expected created_at to be populated by default")
 	}
 }
+
+// ── parseVectorDim ──────────────────────────────────────────────────────────
+
+func TestParseVectorDim(t *testing.T) {
+	cases := []struct {
+		input string
+		dim   int
+		ok    bool
+	}{
+		{"vector(384)", 384, true},
+		{"vector(1)", 1, true},
+		{"vector(4096)", 4096, true},
+		{"Vector(128)", 128, true}, // case insensitive
+		{"vector(0)", 0, false},
+		{"vector(5000)", 0, false},   // exceeds 4096
+		{"vector()", 0, false},
+		{"vector(abc)", 0, false},
+		{"text", 0, false},
+		{"vector", 0, false},
+	}
+	for _, c := range cases {
+		dim, ok := parseVectorDim(c.input)
+		if ok != c.ok || dim != c.dim {
+			t.Errorf("parseVectorDim(%q) = (%d, %v), want (%d, %v)", c.input, dim, ok, c.dim, c.ok)
+		}
+	}
+}
+
+// ── mapColType vector ───────────────────────────────────────────────────────
+
+func TestMapColType_VectorSQLite(t *testing.T) {
+	got := mapColType("vector(384)", false, false)
+	if got != "TEXT" {
+		t.Errorf("vector on SQLite: got %q, want TEXT", got)
+	}
+}
+
+func TestMapColType_VectorPgNoPgvector(t *testing.T) {
+	got := mapColType("vector(384)", true, false)
+	if got != "JSONB" {
+		t.Errorf("vector on PG without pgvector: got %q, want JSONB", got)
+	}
+}
+
+func TestMapColType_VectorPgvector(t *testing.T) {
+	got := mapColType("vector(384)", true, true)
+	if got != "vector(384)" {
+		t.Errorf("vector on PG with pgvector: got %q, want vector(384)", got)
+	}
+}
+
+// ── CreateExtTables with vector column ──────────────────────────────────────
+
+func TestCreateExtTables_VectorColumn(t *testing.T) {
+	db := newSchemaTestDB(t)
+	m := newMemExtDataStore()
+	ctx := context.Background()
+
+	tables := map[string]TableDef{
+		"docs": {
+			Columns: map[string]string{"title": "text", "embedding": "vector(384)"},
+		},
+	}
+	if err := CreateExtTables(ctx, db, false, storesWithExtData(m), "vec-pkg", tables, false); err != nil {
+		t.Fatalf("CreateExtTables: %v", err)
+	}
+
+	// Verify table exists and embedding is TEXT (SQLite fallback).
+	_, err := db.ExecContext(ctx,
+		`INSERT INTO ext_vec_pkg_docs (id, title, embedding) VALUES ('1', 'test', '[0.1,0.2]')`)
+	if err != nil {
+		t.Errorf("table not created or wrong schema: %v", err)
+	}
+}

server/handlers/extensions.go

@@ -17,13 +17,19 @@ import (
 
 // ExtensionHandler serves extension management endpoints.
 type ExtensionHandler struct {
-	stores store.Stores
+	stores       store.Stores
+	capabilities map[string]bool
 }
 
 func NewExtensionHandler(stores store.Stores) *ExtensionHandler {
 	return &ExtensionHandler{stores: stores}
 }
 
+// SetCapabilities stores the detected environment capabilities map.
+func (h *ExtensionHandler) SetCapabilities(caps map[string]bool) {
+	h.capabilities = caps
+}
+
 // validTiers is the set of accepted extension tier values.
 var validTiers = map[string]bool{
 	models.ExtTierBrowser:  true,
@@ -231,7 +237,7 @@ func (h *ExtensionHandler) AdminInstallExtension(c *gin.Context) {
 	SyncManifestPermissions(c, h.stores, pkg.ID, manifestMap)
 
 	if tables, ok := ParseDBTables(manifestMap); ok {
-		if err := CreateExtTables(c.Request.Context(), database.DB, database.IsPostgres(), h.stores, pkg.ID, tables); err != nil {
+		if err := CreateExtTables(c.Request.Context(), database.DB, database.IsPostgres(), h.stores, pkg.ID, tables, h.capabilities["pgvector"]); err != nil {
 			log.Printf("[ext-db] schema create failed for %s: %v", pkg.ID, err)
 		}
 	}

server/handlers/packages.go

@@ -553,7 +553,7 @@ func (h *PackageHandler) registerPackage(c *gin.Context, pkgID string, mInfo *Ma
 // and runs schema migrations.
 func (h *PackageHandler) applySchemaAndPermissions(c *gin.Context, pkgID string, mInfo *ManifestInfo, manifest map[string]any, existing *store.PackageRegistration) error {
 	if tables, ok := ParseDBTables(manifest); ok {
-		if err := CreateExtTables(c.Request.Context(), database.DB, database.IsPostgres(), h.stores, pkgID, tables); err != nil {
+		if err := CreateExtTables(c.Request.Context(), database.DB, database.IsPostgres(), h.stores, pkgID, tables, h.capabilities["pgvector"]); err != nil {
 			log.Printf("[ext-db] schema create failed for %s: %v", pkgID, err)
 		}
 	}
@@ -1047,7 +1047,7 @@ func (h *PackageHandler) UpdatePackage(c *gin.Context) {
 	// 6. Additive schema migration
 	var schemaChanges []string
 	if tables, ok := ParseDBTables(manifest); ok {
-		changes, err := MigrateExtTables(ctx, database.DB, database.IsPostgres(), h.stores, pkgID, tables)
+		changes, err := MigrateExtTables(ctx, database.DB, database.IsPostgres(), h.stores, pkgID, tables, h.capabilities["pgvector"])
 		if err != nil {
 			c.JSON(http.StatusInternalServerError, gin.H{"error": "schema migration failed: " + err.Error()})
 			return

server/handlers/packages_bundled.go

@@ -196,7 +196,7 @@ func installBundledPackage(pkgPath, packagesDir string, stores store.Stores, run
 
 	// Create namespaced DB tables declared in the manifest
 	if tables, ok := ParseDBTables(manifest); ok {
-		if err := CreateExtTables(ctx, database.DB, database.IsPostgres(), stores, pkgID, tables); err != nil {
+		if err := CreateExtTables(ctx, database.DB, database.IsPostgres(), stores, pkgID, tables, detectedCaps["pgvector"]); err != nil {
 			log.Printf("[bundled] [ext-db] schema create failed for %s: %v", pkgID, err)
 		}
 	}

server/handlers/packages_update_test.go

@@ -267,7 +267,7 @@ func TestUpdatePackage_SchemaAddColumn(t *testing.T) {
 			},
 		},
 	})
-	CreateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "schema-pkg", tables)
+	CreateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "schema-pkg", tables, false)
 
 	// Insert a row to verify data preservation
 	physical := extPhysicalTable("schema-pkg", "items")

server/handlers/upgrade_test.go

@@ -42,7 +42,7 @@ func TestUpgrade_SchemaAddIndex(t *testing.T) {
 			},
 		},
 	})
-	CreateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "idx-pkg", tables)
+	CreateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "idx-pkg", tables, false)
 
 	// Insert a row
 	physical := extPhysicalTable("idx-pkg", "events")
@@ -116,7 +116,7 @@ func TestUpgrade_SchemaAddColumnIdempotent(t *testing.T) {
 			},
 		},
 	})
-	CreateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "idem-pkg", tables)
+	CreateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "idem-pkg", tables, false)
 
 	// Insert a row
 	physical := extPhysicalTable("idem-pkg", "items")
@@ -183,7 +183,7 @@ func TestUpgrade_SchemaMultiColumnAdd(t *testing.T) {
 			},
 		},
 	})
-	CreateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "multi-pkg", tables)
+	CreateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "multi-pkg", tables, false)
 
 	// Insert a row
 	physical := extPhysicalTable("multi-pkg", "records")
@@ -521,7 +521,7 @@ func TestUpgrade_MigrateExtTables_NewTable(t *testing.T) {
 		},
 	}
 
-	changes, err := MigrateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "migrate-new", newTables)
+	changes, err := MigrateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "migrate-new", newTables, false)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -559,7 +559,7 @@ func TestUpgrade_MigrateExtTables_AddColumnPreservesRows(t *testing.T) {
 			},
 		},
 	})
-	CreateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "migrate-col", tables)
+	CreateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "migrate-col", tables, false)
 
 	physical := extPhysicalTable("migrate-col", "records")
 	_, err := database.TestDB.ExecContext(ctx,
@@ -574,7 +574,7 @@ func TestUpgrade_MigrateExtTables_AddColumnPreservesRows(t *testing.T) {
 			Columns: map[string]string{"title": "text", "status": "text"},
 		},
 	}
-	changes, err := MigrateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "migrate-col", newTables)
+	changes, err := MigrateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "migrate-col", newTables, false)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -626,7 +626,7 @@ func TestUpgrade_MigrateExtTables_IndexIdempotent(t *testing.T) {
 			},
 		},
 	})
-	CreateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "migrate-idx", tables)
+	CreateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "migrate-idx", tables, false)
 
 	// Migrate with same index — should not fail
 	newTables := map[string]TableDef{
@@ -635,13 +635,13 @@ func TestUpgrade_MigrateExtTables_IndexIdempotent(t *testing.T) {
 			Indexes: [][]string{{"category"}},
 		},
 	}
-	_, err := MigrateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "migrate-idx", newTables)
+	_, err := MigrateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "migrate-idx", newTables, false)
 	if err != nil {
 		t.Fatal("idempotent index migration should not fail:", err)
 	}
 
 	// Run again — still no error
-	_, err = MigrateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "migrate-idx", newTables)
+	_, err = MigrateExtTables(ctx, database.TestDB, database.IsPostgres(), stores, "migrate-idx", newTables, false)
 	if err != nil {
 		t.Fatal("second idempotent index migration should not fail:", err)
 	}

server/main.go

@@ -809,6 +809,7 @@ func main() {
 
 			// Extensions (admin)
 			extAdm := handlers.NewExtensionHandler(stores)
+			extAdm.SetCapabilities(detectedCaps)
 			admin.GET("/extensions", extAdm.AdminListExtensions)
 			admin.POST("/extensions", extAdm.AdminInstallExtension)
 			admin.PUT("/extensions/:id", extAdm.AdminUpdateExtension)

server/sandbox/db_module.go

@@ -30,7 +30,10 @@ import (
 	"crypto/rand"
 	"database/sql"
 	"encoding/hex"
+	"encoding/json"
 	"fmt"
+	"math"
+	"sort"
 	"strings"
 
 	"go.starlark.net/starlark"
@@ -39,10 +42,11 @@ import (
 
 // DBModuleConfig holds configuration for a db module instance.
 type DBModuleConfig struct {
-	PackageID  string
+	PackageID   string
-	CanWrite   bool // true when db.write permission is granted
+	CanWrite    bool // true when db.write permission is granted
-	DB         *sql.DB
+	DB          *sql.DB
-	IsPostgres bool
+	IsPostgres  bool
+	HasPgvector bool // true when pgvector extension is available on Postgres
 }
 
 // allowedViews is the set of platform views extensions may query via db.view().
@@ -55,12 +59,13 @@ var allowedViews = map[string]bool{
 // prefixed with ext_{packageID}_. Write operations are guarded by CanWrite.
 func BuildDBModule(ctx context.Context, cfg DBModuleConfig) *starlarkstruct.Module {
 	fns := starlark.StringDict{
-		"query":       starlark.NewBuiltin("db.query", dbQuery(ctx, cfg)),
+		"query":         starlark.NewBuiltin("db.query", dbQuery(ctx, cfg)),
-		"count":       starlark.NewBuiltin("db.count", dbCount(ctx, cfg)),
+		"count":         starlark.NewBuiltin("db.count", dbCount(ctx, cfg)),
-		"aggregate":   starlark.NewBuiltin("db.aggregate", dbAggregate(ctx, cfg)),
+		"aggregate":     starlark.NewBuiltin("db.aggregate", dbAggregate(ctx, cfg)),
-		"query_batch": starlark.NewBuiltin("db.query_batch", dbQueryBatch(ctx, cfg)),
+		"query_batch":   starlark.NewBuiltin("db.query_batch", dbQueryBatch(ctx, cfg)),
-		"view":        starlark.NewBuiltin("db.view", dbView(ctx, cfg)),
+		"query_similar": starlark.NewBuiltin("db.query_similar", dbQuerySimilar(ctx, cfg)),
-		"list_tables": starlark.NewBuiltin("db.list_tables", dbListTables(ctx, cfg)),
+		"view":          starlark.NewBuiltin("db.view", dbView(ctx, cfg)),
+		"list_tables":   starlark.NewBuiltin("db.list_tables", dbListTables(ctx, cfg)),
 	}
 
 	if cfg.CanWrite {
@@ -160,6 +165,20 @@ func starlarkToGoValue(v starlark.Value) (any, error) {
 		return bool(val), nil
 	case starlark.NoneType:
 		return nil, nil
+	case *starlark.List:
+		goSlice := make([]any, val.Len())
+		for i := 0; i < val.Len(); i++ {
+			elem, err := starlarkToGoValue(val.Index(i))
+			if err != nil {
+				return nil, fmt.Errorf("list[%d]: %w", i, err)
+			}
+			goSlice[i] = elem
+		}
+		b, err := json.Marshal(goSlice)
+		if err != nil {
+			return nil, fmt.Errorf("list marshal: %w", err)
+		}
+		return string(b), nil
 	default:
 		return nil, fmt.Errorf("unsupported type %s", v.Type())
 	}
@@ -926,3 +945,271 @@ func dbDelete(ctx context.Context, cfg DBModuleConfig) func(*starlark.Thread, *s
 		return starlark.True, nil
 	}
 }
+
+// ── Vector Similarity ─────────────────────────
+
+// starlarkListToFloats converts a Starlark list to a Go float64 slice.
+func starlarkListToFloats(list *starlark.List) ([]float64, error) {
+	n := list.Len()
+	if n == 0 {
+		return nil, fmt.Errorf("vector must not be empty")
+	}
+	out := make([]float64, n)
+	for i := 0; i < n; i++ {
+		switch v := list.Index(i).(type) {
+		case starlark.Float:
+			out[i] = float64(v)
+		case starlark.Int:
+			i64, _ := v.Int64()
+			out[i] = float64(i64)
+		default:
+			return nil, fmt.Errorf("vector[%d]: expected number, got %s", i, list.Index(i).Type())
+		}
+	}
+	return out, nil
+}
+
+// floatsToVectorString serializes a float64 slice as a JSON array string.
+// e.g. [0.1, 0.2, 0.3] → "[0.1,0.2,0.3]"
+func floatsToVectorString(v []float64) string {
+	b, _ := json.Marshal(v)
+	return string(b)
+}
+
+// parseVectorJSON parses a JSON array string (or pgvector text) into float64 slice.
+func parseVectorJSON(s string) ([]float64, error) {
+	var out []float64
+	if err := json.Unmarshal([]byte(s), &out); err != nil {
+		return nil, fmt.Errorf("parse vector: %w", err)
+	}
+	return out, nil
+}
+
+// cosineDistance computes cosine distance between two vectors.
+// Returns 0.0 for identical vectors, 1.0 for orthogonal, 2.0 for opposite.
+func cosineDistance(a, b []float64) float64 {
+	if len(a) != len(b) || len(a) == 0 {
+		return 1.0
+	}
+	var dot, normA, normB float64
+	for i := range a {
+		dot += a[i] * b[i]
+		normA += a[i] * a[i]
+		normB += b[i] * b[i]
+	}
+	if normA == 0 || normB == 0 {
+		return 1.0
+	}
+	return 1.0 - (dot / (math.Sqrt(normA) * math.Sqrt(normB)))
+}
+
+// dbQuerySimilar implements db.query_similar(table, column, vector=[], limit=10, filters=None, metric="cosine").
+// Returns rows ordered by distance with an injected _distance float.
+//
+// Three dispatch paths:
+//   - pgvector: native <=> operator with HNSW index
+//   - fallback: fetch rows, compute cosine distance in Go, sort, return top N
+func dbQuerySimilar(ctx context.Context, cfg DBModuleConfig) func(*starlark.Thread, *starlark.Builtin, starlark.Tuple, []starlark.Tuple) (starlark.Value, error) {
+	return func(thread *starlark.Thread, b *starlark.Builtin, args starlark.Tuple, kwargs []starlark.Tuple) (starlark.Value, error) {
+		var table, column string
+		var vectorVal *starlark.List
+		var limit starlark.Int = starlark.MakeInt(10)
+		var filters starlark.Value = starlark.None
+		var metric starlark.String = "cosine"
+
+		if err := starlark.UnpackArgs(b.Name(), args, kwargs,
+			"table", &table,
+			"column", &column,
+			"vector", &vectorVal,
+			"limit?", &limit,
+			"filters?", &filters,
+			"metric?", &metric,
+		); err != nil {
+			return nil, err
+		}
+
+		if string(metric) != "cosine" {
+			return nil, fmt.Errorf("db.query_similar: unsupported metric %q (only \"cosine\" is supported)", string(metric))
+		}
+
+		queryVec, err := starlarkListToFloats(vectorVal)
+		if err != nil {
+			return nil, fmt.Errorf("db.query_similar: %w", err)
+		}
+
+		lim, ok := limit.Int64()
+		if !ok || lim < 1 {
+			lim = 10
+		}
+		if lim > 100 {
+			lim = 100
+		}
+
+		if strings.ContainsAny(column, " \t\n\"';-") {
+			return nil, fmt.Errorf("db.query_similar: invalid column name %q", column)
+		}
+
+		physTable, err := cfg.physicalTable(table)
+		if err != nil {
+			return nil, err
+		}
+
+		if cfg.HasPgvector {
+			return querySimilarPgvector(ctx, cfg, physTable, column, queryVec, int(lim), filters)
+		}
+		return querySimilarFallback(ctx, cfg, physTable, column, queryVec, int(lim), filters)
+	}
+}
+
+// querySimilarPgvector uses the native pgvector <=> operator.
+func querySimilarPgvector(ctx context.Context, cfg DBModuleConfig, physTable, column string, queryVec []float64, limit int, filters starlark.Value) (starlark.Value, error) {
+	vecStr := floatsToVectorString(queryVec)
+
+	whereClause, whereArgs, err := cfg.starlarkFiltersToSQL(filters, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	nextIdx := len(whereArgs) + 1
+	vecPH := cfg.ph(nextIdx)
+	limitPH := cfg.ph(nextIdx + 1)
+
+	query := fmt.Sprintf("SELECT *, (%s <=> %s::vector) AS _distance FROM %s",
+		column, vecPH, physTable)
+	if whereClause != "" {
+		query += " " + whereClause
+	}
+	query += fmt.Sprintf(" ORDER BY %s <=> %s::vector LIMIT %s", column, vecPH, limitPH)
+
+	allArgs := append(whereArgs, vecStr, limit)
+
+	rows, err := cfg.DB.QueryContext(ctx, query, allArgs...)
+	if err != nil {
+		return nil, fmt.Errorf("db.query_similar: %w", err)
+	}
+	defer rows.Close()
+
+	return rowsToStarlark(rows)
+}
+
+// querySimilarFallback fetches rows and computes cosine distance in Go.
+// Used for SQLite (TEXT) and Postgres without pgvector (JSONB).
+func querySimilarFallback(ctx context.Context, cfg DBModuleConfig, physTable, column string, queryVec []float64, limit int, filters starlark.Value) (starlark.Value, error) {
+	whereClause, whereArgs, err := cfg.starlarkFiltersToSQL(filters, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	// Fetch up to 1000 rows for distance computation.
+	fetchLimit := limit * 10
+	if fetchLimit > 1000 {
+		fetchLimit = 1000
+	}
+	if fetchLimit < limit {
+		fetchLimit = limit
+	}
+
+	limitPH := cfg.ph(len(whereArgs) + 1)
+	query := fmt.Sprintf("SELECT * FROM %s", physTable)
+	if whereClause != "" {
+		query += " " + whereClause
+	}
+	query += " LIMIT " + limitPH
+	allArgs := append(whereArgs, fetchLimit)
+
+	rows, err := cfg.DB.QueryContext(ctx, query, allArgs...)
+	if err != nil {
+		return nil, fmt.Errorf("db.query_similar: %w", err)
+	}
+	defer rows.Close()
+
+	cols, err := rows.Columns()
+	if err != nil {
+		return nil, err
+	}
+
+	// Find the vector column index.
+	vecColIdx := -1
+	for i, c := range cols {
+		if c == column {
+			vecColIdx = i
+			break
+		}
+	}
+	if vecColIdx == -1 {
+		return nil, fmt.Errorf("db.query_similar: column %q not found in table", column)
+	}
+
+	type scoredRow struct {
+		vals     []any
+		distance float64
+	}
+	var scored []scoredRow
+
+	for rows.Next() {
+		vals := make([]any, len(cols))
+		ptrs := make([]any, len(cols))
+		for i := range vals {
+			ptrs[i] = &vals[i]
+		}
+		if err := rows.Scan(ptrs...); err != nil {
+			return nil, err
+		}
+
+		// Parse the vector column value.
+		var vecStr string
+		switch v := vals[vecColIdx].(type) {
+		case string:
+			vecStr = v
+		case []byte:
+			vecStr = string(v)
+		default:
+			continue // skip rows with unparseable vectors
+		}
+
+		rowVec, err := parseVectorJSON(vecStr)
+		if err != nil {
+			continue // skip malformed vectors
+		}
+
+		dist := cosineDistance(queryVec, rowVec)
+		scored = append(scored, scoredRow{vals: vals, distance: dist})
+	}
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+
+	// Sort by distance ascending.
+	sort.Slice(scored, func(i, j int) bool {
+		return scored[i].distance < scored[j].distance
+	})
+
+	// Take top N and build Starlark dicts.
+	if len(scored) > limit {
+		scored = scored[:limit]
+	}
+
+	var result []starlark.Value
+	for _, sr := range scored {
+		d := starlark.NewDict(len(cols) + 1)
+		for i, col := range cols {
+			sv, err := goToStarlark(sr.vals[i])
+			if err != nil {
+				return nil, fmt.Errorf("column %q: %w", col, err)
+			}
+			if err := d.SetKey(starlark.String(col), sv); err != nil {
+				return nil, err
+			}
+		}
+		// Inject _distance.
+		if err := d.SetKey(starlark.String("_distance"), starlark.Float(sr.distance)); err != nil {
+			return nil, err
+		}
+		result = append(result, d)
+	}
+
+	if result == nil {
+		result = []starlark.Value{}
+	}
+	return starlark.NewList(result), nil
+}

server/sandbox/db_module_test.go

@@ -3,6 +3,7 @@ package sandbox
 import (
 	"context"
 	"database/sql"
+	"math"
 	"strings"
 	"testing"
 
@@ -684,3 +685,252 @@ func TestDBQueryBatch_MissingTable(t *testing.T) {
 	}
 }
 
+// ── starlarkToGoValue list ──────────────────
+
+func TestStarlarkToGoValue_List(t *testing.T) {
+	list := starlark.NewList([]starlark.Value{
+		starlark.Float(0.1), starlark.Float(0.2), starlark.Float(0.3),
+	})
+	got, err := starlarkToGoValue(list)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	s, ok := got.(string)
+	if !ok {
+		t.Fatalf("expected string, got %T", got)
+	}
+	if s != "[0.1,0.2,0.3]" {
+		t.Errorf("got %q, want %q", s, "[0.1,0.2,0.3]")
+	}
+}
+
+// ── cosineDistance ───────────────────────────
+
+func TestCosineDistance(t *testing.T) {
+	// Identical vectors → distance 0
+	d := cosineDistance([]float64{1, 0, 0}, []float64{1, 0, 0})
+	if math.Abs(d) > 1e-9 {
+		t.Errorf("identical vectors: got %f, want 0", d)
+	}
+
+	// Orthogonal vectors → distance 1
+	d = cosineDistance([]float64{1, 0}, []float64{0, 1})
+	if math.Abs(d-1.0) > 1e-9 {
+		t.Errorf("orthogonal vectors: got %f, want 1", d)
+	}
+
+	// Opposite vectors → distance 2
+	d = cosineDistance([]float64{1, 0}, []float64{-1, 0})
+	if math.Abs(d-2.0) > 1e-9 {
+		t.Errorf("opposite vectors: got %f, want 2", d)
+	}
+
+	// Empty/mismatched → distance 1
+	d = cosineDistance([]float64{}, []float64{})
+	if d != 1.0 {
+		t.Errorf("empty vectors: got %f, want 1", d)
+	}
+	d = cosineDistance([]float64{1}, []float64{1, 2})
+	if d != 1.0 {
+		t.Errorf("mismatched vectors: got %f, want 1", d)
+	}
+}
+
+// ── vector helper: newVectorTestDB ──────────
+
+func newVectorTestDB(t *testing.T) (*sql.DB, DBModuleConfig) {
+	t.Helper()
+	db, err := sql.Open("sqlite", ":memory:")
+	if err != nil {
+		t.Fatalf("open db: %v", err)
+	}
+	t.Cleanup(func() { db.Close() })
+
+	// Create a table with a TEXT column for vector storage (SQLite fallback).
+	_, err = db.Exec(`CREATE TABLE ext_test_ext_embeddings (
+		id         TEXT PRIMARY KEY,
+		embedding  TEXT,
+		category   TEXT,
+		created_at TEXT DEFAULT (datetime('now'))
+	)`)
+	if err != nil {
+		t.Fatalf("create test table: %v", err)
+	}
+
+	cfg := DBModuleConfig{
+		PackageID:   "test-ext",
+		CanWrite:    true,
+		DB:          db,
+		IsPostgres:  false,
+		HasPgvector: false,
+	}
+	return db, cfg
+}
+
+// ── db.query_similar ────────────────────────
+
+func TestDBQuerySimilar_Basic(t *testing.T) {
+	db, cfg := newVectorTestDB(t)
+
+	// Insert 5 rows with known vectors.
+	db.Exec(`INSERT INTO ext_test_ext_embeddings (id, embedding, category) VALUES ('a', '[1,0,0]', 'x')`)
+	db.Exec(`INSERT INTO ext_test_ext_embeddings (id, embedding, category) VALUES ('b', '[0,1,0]', 'x')`)
+	db.Exec(`INSERT INTO ext_test_ext_embeddings (id, embedding, category) VALUES ('c', '[0,0,1]', 'x')`)
+	db.Exec(`INSERT INTO ext_test_ext_embeddings (id, embedding, category) VALUES ('d', '[0.9,0.1,0]', 'x')`)
+	db.Exec(`INSERT INTO ext_test_ext_embeddings (id, embedding, category) VALUES ('e', '[0,0.9,0.1]', 'x')`)
+
+	// Query similar to [1,0,0] — should return 'a' first (identical), then 'd' (close).
+	result, err := execScript(t, cfg, `
+rows = db.query_similar("embeddings", "embedding", vector=[1.0, 0.0, 0.0], limit=3)
+`)
+	if err != nil {
+		t.Fatalf("script error: %v", err)
+	}
+
+	rows, ok := result.Globals["rows"].(*starlark.List)
+	if !ok {
+		t.Fatalf("rows is %T, want *starlark.List", result.Globals["rows"])
+	}
+	if rows.Len() != 3 {
+		t.Fatalf("got %d rows, want 3", rows.Len())
+	}
+
+	// First row should be 'a' (distance ≈ 0).
+	first := rows.Index(0).(*starlark.Dict)
+	idVal, _, _ := first.Get(starlark.String("id"))
+	if string(idVal.(starlark.String)) != "a" {
+		t.Errorf("first row id = %s, want 'a'", idVal)
+	}
+
+	// Verify _distance is present and near 0.
+	distVal, _, _ := first.Get(starlark.String("_distance"))
+	dist := float64(distVal.(starlark.Float))
+	if dist > 0.01 {
+		t.Errorf("first row _distance = %f, want ≈ 0", dist)
+	}
+
+	// Second row should be 'd' (closest after identical).
+	second := rows.Index(1).(*starlark.Dict)
+	idVal2, _, _ := second.Get(starlark.String("id"))
+	if string(idVal2.(starlark.String)) != "d" {
+		t.Errorf("second row id = %s, want 'd'", idVal2)
+	}
+}
+
+func TestDBQuerySimilar_WithFilters(t *testing.T) {
+	db, cfg := newVectorTestDB(t)
+
+	db.Exec(`INSERT INTO ext_test_ext_embeddings (id, embedding, category) VALUES ('a', '[1,0,0]', 'alpha')`)
+	db.Exec(`INSERT INTO ext_test_ext_embeddings (id, embedding, category) VALUES ('b', '[0.9,0.1,0]', 'beta')`)
+	db.Exec(`INSERT INTO ext_test_ext_embeddings (id, embedding, category) VALUES ('c', '[0,1,0]', 'alpha')`)
+
+	// Filter to alpha only — should exclude 'b' even though it's close.
+	result, err := execScript(t, cfg, `
+rows = db.query_similar("embeddings", "embedding", vector=[1.0, 0.0, 0.0], filters={"category": "alpha"})
+`)
+	if err != nil {
+		t.Fatalf("script error: %v", err)
+	}
+
+	rows := result.Globals["rows"].(*starlark.List)
+	for i := 0; i < rows.Len(); i++ {
+		d := rows.Index(i).(*starlark.Dict)
+		idVal, _, _ := d.Get(starlark.String("id"))
+		if string(idVal.(starlark.String)) == "b" {
+			t.Error("filtered query should not include row 'b' (category=beta)")
+		}
+	}
+}
+
+func TestDBQuerySimilar_EmptyTable(t *testing.T) {
+	_, cfg := newVectorTestDB(t)
+
+	result, err := execScript(t, cfg, `
+rows = db.query_similar("embeddings", "embedding", vector=[1.0, 0.0, 0.0])
+`)
+	if err != nil {
+		t.Fatalf("script error: %v", err)
+	}
+	rows := result.Globals["rows"].(*starlark.List)
+	if rows.Len() != 0 {
+		t.Errorf("expected 0 rows for empty table, got %d", rows.Len())
+	}
+}
+
+func TestDBQuerySimilar_InvalidMetric(t *testing.T) {
+	_, cfg := newVectorTestDB(t)
+
+	_, err := execScript(t, cfg, `
+rows = db.query_similar("embeddings", "embedding", vector=[1.0], metric="l2")
+`)
+	if err == nil {
+		t.Fatal("expected error for unsupported metric")
+	}
+	if !strings.Contains(err.Error(), "unsupported metric") {
+		t.Errorf("unexpected error: %v", err)
+	}
+}
+
+func TestDBQuerySimilar_LimitCap(t *testing.T) {
+	db, cfg := newVectorTestDB(t)
+
+	// Insert 5 rows
+	for i := 0; i < 5; i++ {
+		db.Exec(`INSERT INTO ext_test_ext_embeddings (id, embedding) VALUES (?, '[1,0,0]')`,
+			generateID())
+	}
+
+	result, err := execScript(t, cfg, `
+rows = db.query_similar("embeddings", "embedding", vector=[1.0, 0.0, 0.0], limit=2)
+`)
+	if err != nil {
+		t.Fatalf("script error: %v", err)
+	}
+	rows := result.Globals["rows"].(*starlark.List)
+	if rows.Len() != 2 {
+		t.Errorf("expected 2 rows with limit=2, got %d", rows.Len())
+	}
+}
+
+func TestDBInsert_VectorColumn(t *testing.T) {
+	db, cfg := newVectorTestDB(t)
+
+	_, err := execScript(t, cfg, `
+row = db.insert("embeddings", {"embedding": [0.1, 0.2, 0.3], "category": "test"})
+`)
+	if err != nil {
+		t.Fatalf("script error: %v", err)
+	}
+
+	// Verify the stored value is valid JSON.
+	var stored string
+	db.QueryRow(`SELECT embedding FROM ext_test_ext_embeddings WHERE category = 'test'`).Scan(&stored)
+	if stored != "[0.1,0.2,0.3]" {
+		t.Errorf("stored vector = %q, want %q", stored, "[0.1,0.2,0.3]")
+	}
+}
+
+func TestDBQuerySimilar_InsertThenQuery(t *testing.T) {
+	_, cfg := newVectorTestDB(t)
+
+	// Full round-trip: insert via db.insert, then query_similar.
+	result, err := execScript(t, cfg, `
+db.insert("embeddings", {"embedding": [1.0, 0.0, 0.0], "category": "a"})
+db.insert("embeddings", {"embedding": [0.0, 1.0, 0.0], "category": "b"})
+rows = db.query_similar("embeddings", "embedding", vector=[1.0, 0.0, 0.0], limit=1)
+`)
+	if err != nil {
+		t.Fatalf("script error: %v", err)
+	}
+
+	rows := result.Globals["rows"].(*starlark.List)
+	if rows.Len() != 1 {
+		t.Fatalf("expected 1 row, got %d", rows.Len())
+	}
+	first := rows.Index(0).(*starlark.Dict)
+	catVal, _, _ := first.Get(starlark.String("category"))
+	if string(catVal.(starlark.String)) != "a" {
+		t.Errorf("closest match should be category=a, got %s", catVal)
+	}
+}
+

server/sandbox/runner.go

@@ -425,10 +425,11 @@ func (r *Runner) buildModulesWithLibCtx(ctx context.Context, packageID string, m
 	// Wire db module at the highest granted level.
 	if dbLevel > 0 && r.db != nil {
 		modules["db"] = BuildDBModule(ctx, DBModuleConfig{
-			PackageID:  packageID,
+			PackageID:   packageID,
-			CanWrite:   dbLevel == 2,
+			CanWrite:    dbLevel == 2,
-			DB:         r.db,
+			DB:          r.db,
-			IsPostgres: r.dbPostgres,
+			IsPostgres:  r.dbPostgres,
+			HasPgvector: r.capabilities["pgvector"],
 		})
 	}