@gmickel/gno 0.38.0 → 0.39.0

package/README.md

@@ -92,7 +92,7 @@ gno daemon
 
 - **Retrieval Quality Upgrade**: stronger BM25 lexical handling, code-aware chunking, terminal result hyperlinks, and per-collection model overrides
 - **Code Embedding Benchmarks**: new benchmark workflow across canonical, real-GNO, and pinned OSS slices for comparing alternate embedding models
-- **Recommended Code Embed Model**: docs and benchmark pages now point to `Qwen3-Embedding-0.6B-GGUF` as the current code-specialist embedding option
+- **Default Embed Model**: built-in presets now use `Qwen3-Embedding-0.6B-GGUF` after it beat `bge-m3` on both code and multilingual prose benchmark lanes
 - **Regression Fixes**: tightened phrase/negation/hyphen/underscore BM25 behavior, cleaned non-TTY hyperlink output, improved `gno doctor` chunking visibility, and fixed the embedding autoresearch harness
 
 ### Fine-Tuned Model Quick Use
@@ -103,7 +103,7 @@ models:
   presets:
     - id: slim-tuned
       name: GNO Slim Tuned
-      embed: hf:gpustack/bge-m3-GGUF/bge-m3-Q4_K_M.gguf
+      embed: hf:Qwen/Qwen3-Embedding-0.6B-GGUF/Qwen3-Embedding-0.6B-Q8_0.gguf
       rerank: hf:ggml-org/Qwen3-Reranker-0.6B-Q8_0-GGUF/qwen3-reranker-0.6b-q8_0.gguf
       expand: hf:guiltylemon/gno-expansion-slim-retrieval-v1/gno-expansion-auto-entity-lock-default-mix-lr95-f16.gguf
       gen: hf:unsloth/Qwen3-1.7B-GGUF/Qwen3-1.7B-Q4_K_M.gguf
@@ -659,11 +659,11 @@ graph TD
 
 Models auto-download on first use to `~/.cache/gno/models/`. For deterministic startup, set `GNO_NO_AUTO_DOWNLOAD=1` and use `gno models pull` explicitly. Alternatively, offload to a GPU server on your network using HTTP backends.
 
-| Model               | Purpose                               | Size         |
-| :------------------ | :------------------------------------ | :----------- |
-| bge-m3              | Embeddings (1024-dim, multilingual)   | ~500MB       |
-| Qwen3-Reranker-0.6B | Cross-encoder reranking (32K context) | ~700MB       |
-| Qwen/SmolLM         | Query expansion + AI answers          | ~600MB-1.2GB |
+| Model                | Purpose                               | Size         |
+| :------------------- | :------------------------------------ | :----------- |
+| Qwen3-Embedding-0.6B | Embeddings (multilingual)             | ~640MB       |
+| Qwen3-Reranker-0.6B  | Cross-encoder reranking (32K context) | ~700MB       |
+| Qwen/SmolLM          | Query expansion + AI answers          | ~600MB-1.2GB |
 
 ### Model Presets
 
@@ -814,7 +814,28 @@ Why this is the current recommendation:
 Trade-off:
 
 - Qwen is slower to embed than `bge-m3`
-- use it where code retrieval quality matters, not necessarily as the global default for every collection
+- existing users upgrading to the new default may need to run `gno embed` again so vector and hybrid retrieval catch up
+
+### General Multilingual Embedding Benchmark
+
+GNO also now has a separate public-docs benchmark lane for normal markdown/prose
+collections:
+
+```bash
+bun run bench:general-embeddings --candidate bge-m3-incumbent --write
+bun run bench:general-embeddings --candidate qwen3-embedding-0.6b --write
+```
+
+Current signal on the public multilingual FastAPI-docs fixture:
+
+- `bge-m3`: vector nDCG@10 `0.350`, hybrid nDCG@10 `0.642`
+- `Qwen3-Embedding-0.6B-GGUF`: vector nDCG@10 `0.859`, hybrid nDCG@10 `0.947`
+
+Interpretation:
+
+- Qwen is now the strongest general multilingual embedding model we have tested
+- built-in presets now use Qwen by default
+- existing users may need to run `gno embed` again after upgrading so current collections catch up
 
 ---
 

package/assets/skill/SKILL.md

@@ -199,6 +199,13 @@ Good default guidance:
 gno embed --collection gno-code
 ```
 
+If you want to remove old vectors after switching:
+
+```bash
+gno collection clear-embeddings gno-code        # stale models only
+gno collection clear-embeddings gno-code --all  # remove everything, then re-embed
+```
+
 ## Reference Documentation
 
 | Topic                                                 | File                                 |

package/assets/skill/cli-reference.md

@@ -66,6 +66,12 @@ gno collection remove <name>
 gno collection rename <old> <new>
 ```
 
+### gno collection clear-embeddings
+
+```bash
+gno collection clear-embeddings <name> [--all] [--json]
+```
+
 ## Indexing
 
 ### gno update

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gmickel/gno",
-  "version": "0.38.0",
+  "version": "0.39.0",
   "description": "Local semantic search for your documents. Index Markdown, PDF, and Office files with hybrid BM25 + vector search.",
   "keywords": [
     "embeddings",
@@ -71,6 +71,8 @@
     "eval:hybrid:delta": "bun scripts/hybrid-benchmark.ts --delta",
     "bench:code-embeddings": "bun scripts/code-embedding-benchmark.ts",
     "bench:code-embeddings:write": "bun scripts/code-embedding-benchmark.ts --write",
+    "bench:general-embeddings": "bun scripts/general-embedding-benchmark.ts",
+    "bench:general-embeddings:write": "bun scripts/general-embedding-benchmark.ts --write",
     "eval:retrieval-candidates": "bun scripts/retrieval-candidate-benchmark.ts",
     "eval:retrieval-candidates:write": "bun scripts/retrieval-candidate-benchmark.ts --write",
     "eval:watch": "bun --bun evalite watch",

package/src/cli/commands/collection/clear-embeddings.ts

@@ -0,0 +1,83 @@
+/**
+ * gno collection clear-embeddings - Remove stale or all embeddings for a collection.
+ */
+
+import { getIndexDbPath } from "../../../app/constants";
+import { loadConfig } from "../../../config";
+import { resolveModelUri } from "../../../llm/registry";
+import { SqliteAdapter } from "../../../store/sqlite/adapter";
+import { CliError } from "../../errors";
+
+interface ClearEmbeddingsOptions {
+  all?: boolean;
+  json?: boolean;
+}
+
+export async function collectionClearEmbeddings(
+  name: string,
+  options: ClearEmbeddingsOptions = {}
+): Promise<void> {
+  const configResult = await loadConfig();
+  if (!configResult.ok) {
+    throw new CliError(
+      "RUNTIME",
+      `Failed to load config: ${configResult.error.message}`
+    );
+  }
+
+  const config = configResult.value;
+  const collection = config.collections.find(
+    (item) => item.name === name.toLowerCase()
+  );
+  if (!collection) {
+    throw new CliError("VALIDATION", `Collection not found: ${name}`);
+  }
+
+  const store = new SqliteAdapter();
+  const openResult = await store.open(getIndexDbPath(), config.ftsTokenizer);
+  if (!openResult.ok) {
+    throw new CliError("RUNTIME", openResult.error.message);
+  }
+
+  try {
+    const mode = options.all ? "all" : "stale";
+    const activeModel = resolveModelUri(
+      config,
+      "embed",
+      undefined,
+      collection.name
+    );
+    const result = await store.clearEmbeddingsForCollection(collection.name, {
+      mode,
+      activeModel,
+    });
+
+    if (!result.ok) {
+      throw new CliError("RUNTIME", result.error.message);
+    }
+
+    if (options.json) {
+      process.stdout.write(`${JSON.stringify(result.value, null, 2)}\n`);
+      return;
+    }
+
+    const lines = [
+      `Cleared ${result.value.deletedVectors} embedding(s) for ${result.value.collection}.`,
+      `Mode: ${result.value.mode}`,
+    ];
+    if (result.value.deletedModels.length > 0) {
+      lines.push(`Models: ${result.value.deletedModels.join(", ")}`);
+    }
+    if (result.value.protectedSharedVectors > 0) {
+      lines.push(
+        `Retained ${result.value.protectedSharedVectors} shared vector(s) still referenced by other active collections.`
+      );
+    }
+    if (mode === "all") {
+      lines.push(`Run: gno embed --collection ${result.value.collection}`);
+    }
+    process.stdout.write(`${lines.join("\n")}\n`);
+  } finally {
+    await store.close();
+  }
+}

package/src/cli/commands/collection/index.ts

@@ -5,6 +5,7 @@
  */
 
 export { collectionAdd } from "./add";
+export { collectionClearEmbeddings } from "./clear-embeddings";
 export { collectionList } from "./list";
 export { collectionRemove } from "./remove";
 export { collectionRename } from "./rename";

package/src/cli/commands/models/use.ts

@@ -7,7 +7,7 @@
 
 import { createDefaultConfig, loadConfig } from "../../../config";
 import { saveConfig } from "../../../config/saver";
-import { getPreset, listPresets } from "../../../llm/registry";
+import { getPreset, listPresets, resolveModelUri } from "../../../llm/registry";
 
 // ─────────────────────────────────────────────────────────────────────────────
 // Types
@@ -19,7 +19,12 @@ export interface ModelsUseOptions {
 }
 
 export type ModelsUseResult =
-  | { success: true; preset: string; name: string }
+  | {
+      success: true;
+      preset: string;
+      name: string;
+      embedModelChanged: boolean;
+    }
   | { success: false; error: string };
 
 // ─────────────────────────────────────────────────────────────────────────────
@@ -36,6 +41,7 @@ export async function modelsUse(
   // Load existing config or create default
   const configResult = await loadConfig(options.configPath);
   const config = configResult.ok ? configResult.value : createDefaultConfig();
+  const previousEmbedModel = resolveModelUri(config, "embed");
 
   // Check if preset exists
   const preset = getPreset(config, presetId);
@@ -72,7 +78,12 @@ export async function modelsUse(
     };
   }
 
-  return { success: true, preset: presetId, name: preset.name };
+  return {
+    success: true,
+    preset: presetId,
+    name: preset.name,
+    embedModelChanged: previousEmbedModel !== preset.embed,
+  };
 }
 
 /**
@@ -82,5 +93,9 @@ export function formatModelsUse(result: ModelsUseResult): string {
   if (!result.success) {
     return `Error: ${result.error}`;
   }
-  return `Switched to preset: ${result.preset} (${result.name})`;
+  const lines = [`Switched to preset: ${result.preset} (${result.name})`];
+  if (result.embedModelChanged) {
+    lines.push("Embedding model changed. Run: gno embed");
+  }
+  return lines.join("\n");
 }

package/src/cli/commands/status.ts

@@ -9,6 +9,7 @@ import type { IndexStatus } from "../../store/types";
 
 import { getIndexDbPath } from "../../app/constants";
 import { getConfigPaths, isInitialized, loadConfig } from "../../config";
+import { resolveModelUri } from "../../llm/registry";
 import { SqliteAdapter } from "../../store/sqlite/adapter";
 
 /**
@@ -148,7 +149,9 @@ export async function status(
   }
 
   try {
-    const statusResult = await store.getStatus();
+    const statusResult = await store.getStatus({
+      embedModel: resolveModelUri(config, "embed"),
+    });
     if (!statusResult.ok) {
       return { success: false, error: statusResult.error.message };
     }

package/src/cli/program.ts

@@ -1310,6 +1310,20 @@ function wireManagementCommands(program: Command): void {
       await collectionRename(oldName, newName);
     });
 
+  collectionCmd
+    .command("clear-embeddings <name>")
+    .description("Clear stale or all embeddings for a collection")
+    .option("--all", "remove all embeddings for the collection")
+    .option("--json", "JSON output")
+    .action(async (name: string, cmdOpts: Record<string, unknown>) => {
+      const { collectionClearEmbeddings } =
+        await import("./commands/collection");
+      await collectionClearEmbeddings(name, {
+        all: Boolean(cmdOpts.all),
+        json: Boolean(cmdOpts.json),
+      });
+    });
+
   // context subcommands
   const contextCmd = program
     .command("context")

package/src/config/types.ts

@@ -190,7 +190,7 @@ export const DEFAULT_MODEL_PRESETS: ModelPreset[] = [
   {
     id: "slim-tuned",
     name: "GNO Slim Tuned (Default, ~1GB)",
-    embed: "hf:gpustack/bge-m3-GGUF/bge-m3-Q4_K_M.gguf",
+    embed: "hf:Qwen/Qwen3-Embedding-0.6B-GGUF/Qwen3-Embedding-0.6B-Q8_0.gguf",
     rerank:
       "hf:ggml-org/Qwen3-Reranker-0.6B-Q8_0-GGUF/qwen3-reranker-0.6b-q8_0.gguf",
     expand:
@@ -200,7 +200,7 @@ export const DEFAULT_MODEL_PRESETS: ModelPreset[] = [
   {
     id: "slim",
     name: "Slim (~1GB)",
-    embed: "hf:gpustack/bge-m3-GGUF/bge-m3-Q4_K_M.gguf",
+    embed: "hf:Qwen/Qwen3-Embedding-0.6B-GGUF/Qwen3-Embedding-0.6B-Q8_0.gguf",
     rerank:
       "hf:ggml-org/Qwen3-Reranker-0.6B-Q8_0-GGUF/qwen3-reranker-0.6b-q8_0.gguf",
     expand: "hf:unsloth/Qwen3-1.7B-GGUF/Qwen3-1.7B-Q4_K_M.gguf",
@@ -209,7 +209,7 @@ export const DEFAULT_MODEL_PRESETS: ModelPreset[] = [
   {
     id: "balanced",
     name: "Balanced (~2GB)",
-    embed: "hf:gpustack/bge-m3-GGUF/bge-m3-Q4_K_M.gguf",
+    embed: "hf:Qwen/Qwen3-Embedding-0.6B-GGUF/Qwen3-Embedding-0.6B-Q8_0.gguf",
     rerank:
       "hf:ggml-org/Qwen3-Reranker-0.6B-Q8_0-GGUF/qwen3-reranker-0.6b-q8_0.gguf",
     expand:
@@ -219,7 +219,7 @@ export const DEFAULT_MODEL_PRESETS: ModelPreset[] = [
   {
     id: "quality",
     name: "Quality (Best Answers, ~2.5GB)",
-    embed: "hf:gpustack/bge-m3-GGUF/bge-m3-Q4_K_M.gguf",
+    embed: "hf:Qwen/Qwen3-Embedding-0.6B-GGUF/Qwen3-Embedding-0.6B-Q8_0.gguf",
     rerank:
       "hf:ggml-org/Qwen3-Reranker-0.6B-Q8_0-GGUF/qwen3-reranker-0.6b-q8_0.gguf",
     expand:

package/src/mcp/tools/status.ts

@@ -7,6 +7,7 @@
 import type { IndexStatus } from "../../store/types";
 import type { ToolContext } from "../server";
 
+import { resolveModelUri } from "../../llm/registry";
 import { runTool, type ToolResult } from "./index";
 
 type StatusInput = Record<string, never>;
@@ -66,7 +67,9 @@ export function handleStatus(
     ctx,
     "gno_status",
     async () => {
-      const result = await ctx.store.getStatus();
+      const result = await ctx.store.getStatus({
+        embedModel: resolveModelUri(ctx.config, "embed"),
+      });
       if (!result.ok) {
         throw new Error(result.error.message);
       }

package/src/sdk/client.ts

@@ -621,7 +621,11 @@ class GnoClientImpl implements GnoClient {
 
   async status(): Promise<IndexStatus> {
     this.assertOpen();
-    return unwrapStore(await this.store.getStatus());
+    return unwrapStore(
+      await this.store.getStatus({
+        embedModel: resolveModelUri(this.config, "embed"),
+      })
+    );
   }
 
   async update(options: GnoUpdateOptions = {}): Promise<SyncResult> {

package/src/serve/public/components/AIModelSelector.tsx

@@ -55,6 +55,8 @@ interface SetPresetResponse {
   success: boolean;
   activePreset: string;
   capabilities: Capabilities;
+  embedModelChanged?: boolean;
+  note?: string;
 }
 
 interface DownloadProgress {
@@ -355,7 +357,12 @@ export function AIModelSelector({
       checkCapabilities(data.capabilities);
       setOpen(false);
       const presetName = presets.find((preset) => preset.id === id)?.name ?? id;
-      setNotice(`Switched to ${presetName}`);
+      setNotice(
+        data.embedModelChanged
+          ? (data.note ??
+              `Switched to ${presetName}. Run embeddings again so vector results catch up.`)
+          : `Switched to ${presetName}`
+      );
       const { data: statusData } =
         await apiFetch<AppStatusResponse>("/api/status");
       await syncFromStatus(statusData);

package/src/serve/public/components/CollectionModelDialog.tsx

@@ -133,7 +133,9 @@ export function CollectionModelDialog({
   const [error, setError] = useState<string | null>(null);
   const [saving, setSaving] = useState(false);
   const showCodeRecommendation =
-    collection !== null && collectionLooksCodeHeavy(collection);
+    collection !== null &&
+    collectionLooksCodeHeavy(collection) &&
+    collection.effectiveModels?.embed !== CODE_EMBED_RECOMMENDATION;
 
   useEffect(() => {
     if (!open || !collection) {

package/src/serve/public/pages/Collections.tsx

@@ -100,12 +100,25 @@ interface SyncResponse {
   jobId: string;
 }
 
+interface EmbeddingCleanupResponse {
+  note?: string;
+  stats: {
+    collection: string;
+    deletedVectors: number;
+    deletedModels: string[];
+    mode: "stale" | "all";
+    protectedSharedVectors: number;
+  };
+  success: boolean;
+}
+
 interface CollectionsResponseItem extends CollectionModelDetails {}
 
 interface CollectionCardProps {
   actionsDisabled: boolean;
   collection: CollectionStats;
   onBrowse: () => void;
+  onEmbeddingCleanup: () => void;
   onModelSettings: () => void;
   onReindex: () => void;
   onRemove: () => void;
@@ -135,6 +148,7 @@ function CollectionCard({
   actionsDisabled,
   collection,
   onBrowse,
+  onEmbeddingCleanup,
   onModelSettings,
   onReindex,
   onRemove,
@@ -192,6 +206,16 @@ function CollectionCard({
                 <CpuIcon className="mr-2 size-4" />
                 Model settings
               </DropdownMenuItem>
+              <DropdownMenuItem
+                disabled={actionsDisabled}
+                onClick={(event) => {
+                  event.stopPropagation();
+                  onEmbeddingCleanup();
+                }}
+              >
+                <DatabaseIcon className="mr-2 size-4" />
+                Embedding cleanup
+              </DropdownMenuItem>
               <DropdownMenuSeparator />
               <DropdownMenuItem
                 disabled={actionsDisabled || isReindexing}
@@ -316,6 +340,14 @@ export default function Collections({ navigate }: PageProps) {
   const [removeDialog, setRemoveDialog] = useState<CollectionStats | null>(
     null
   );
+  const [embeddingCleanupDialog, setEmbeddingCleanupDialog] =
+    useState<CollectionStats | null>(null);
+  const [embeddingCleanupBusy, setEmbeddingCleanupBusy] = useState<
+    "stale" | "all" | null
+  >(null);
+  const [embeddingCleanupNote, setEmbeddingCleanupNote] = useState<
+    string | null
+  >(null);
   const [removing, setRemoving] = useState(false);
   const [addDialogOpen, setAddDialogOpen] = useState(false);
   const [modelDialogCollection, setModelDialogCollection] =
@@ -410,6 +442,34 @@ export default function Collections({ navigate }: PageProps) {
     }
   };
 
+  const handleEmbeddingCleanup = async (mode: "stale" | "all") => {
+    if (!embeddingCleanupDialog) return;
+
+    setEmbeddingCleanupBusy(mode);
+    setEmbeddingCleanupNote(null);
+
+    const { data, error: err } = await apiFetch<EmbeddingCleanupResponse>(
+      `/api/collections/${encodeURIComponent(embeddingCleanupDialog.name)}/embeddings/clear`,
+      {
+        method: "POST",
+        body: JSON.stringify({ mode }),
+      }
+    );
+
+    setEmbeddingCleanupBusy(null);
+
+    if (err) {
+      setEmbeddingCleanupNote(err);
+      return;
+    }
+
+    setEmbeddingCleanupNote(
+      data?.note ??
+        `Cleared ${data?.stats.deletedVectors ?? 0} embedding(s) for ${embeddingCleanupDialog.name}.`
+    );
+    await loadCollections();
+  };
+
   // Loading state
   if (loading) {
     return (
@@ -577,6 +637,10 @@ export default function Collections({ navigate }: PageProps) {
                     `/browse?collection=${encodeURIComponent(collection.name)}`
                   )
                 }
+                onEmbeddingCleanup={() => {
+                  setEmbeddingCleanupDialog(collection);
+                  setEmbeddingCleanupNote(null);
+                }}
                 onModelSettings={() => setModelDialogCollection(collection)}
                 onReindex={() => void handleReindex(collection.name)}
                 onRemove={() => setRemoveDialog(collection)}
@@ -605,6 +669,80 @@ export default function Collections({ navigate }: PageProps) {
         open={!!modelDialogCollection}
       />
 
+      <Dialog
+        onOpenChange={(open) => {
+          if (!open) {
+            setEmbeddingCleanupDialog(null);
+            setEmbeddingCleanupBusy(null);
+            setEmbeddingCleanupNote(null);
+          }
+        }}
+        open={!!embeddingCleanupDialog}
+      >
+        <DialogContent>
+          <DialogHeader>
+            <DialogTitle>Embedding cleanup</DialogTitle>
+            <DialogDescription>
+              Remove stale embeddings or clear all embeddings for{" "}
+              <strong>{embeddingCleanupDialog?.name}</strong>.
+            </DialogDescription>
+          </DialogHeader>
+          <div className="space-y-3 text-sm">
+            <div className="rounded-lg border border-border/40 bg-card/70 p-3">
+              <p className="font-medium">Clear stale embeddings</p>
+              <p className="mt-1 text-muted-foreground">
+                Removes embeddings for models that are not the currently active
+                embed model for this collection.
+              </p>
+            </div>
+            <div className="rounded-lg border border-secondary/30 bg-secondary/8 p-3">
+              <p className="font-medium text-secondary">Clear all embeddings</p>
+              <p className="mt-1 text-muted-foreground">
+                Removes every embedding for this collection. You will need to
+                run embeddings again afterward.
+              </p>
+            </div>
+            {embeddingCleanupNote ? (
+              <div className="rounded-lg border border-border/40 bg-card/70 p-3 text-muted-foreground">
+                {embeddingCleanupNote}
+              </div>
+            ) : null}
+          </div>
+          <DialogFooter className="gap-2 sm:gap-0">
+            <Button
+              onClick={() => {
+                setEmbeddingCleanupDialog(null);
+                setEmbeddingCleanupBusy(null);
+                setEmbeddingCleanupNote(null);
+              }}
+              variant="outline"
+            >
+              Cancel
+            </Button>
+            <Button
+              disabled={embeddingCleanupBusy !== null}
+              onClick={() => void handleEmbeddingCleanup("stale")}
+              variant="outline"
+            >
+              {embeddingCleanupBusy === "stale" ? (
+                <Loader2Icon className="mr-1.5 size-4 animate-spin" />
+              ) : null}
+              Clear stale
+            </Button>
+            <Button
+              disabled={embeddingCleanupBusy !== null}
+              onClick={() => void handleEmbeddingCleanup("all")}
+              variant="destructive"
+            >
+              {embeddingCleanupBusy === "all" ? (
+                <Loader2Icon className="mr-1.5 size-4 animate-spin" />
+              ) : null}
+              Clear all
+            </Button>
+          </DialogFooter>
+        </DialogContent>
+      </Dialog>
+
       {/* Remove confirmation dialog */}
       <Dialog
         onOpenChange={(open) => !open && setRemoveDialog(null)}

package/src/serve/routes/api.ts

@@ -82,6 +82,7 @@ import {
   getModelConfig,
   getPreset,
   listPresets,
+  resolveModelUri,
 } from "../../llm/registry";
 import {
   generateGroundedAnswer,
@@ -212,6 +213,10 @@ export interface UpdateCollectionRequestBody {
   };
 }
 
+export interface ClearCollectionEmbeddingsRequestBody {
+  mode?: "stale" | "all";
+}
+
 export interface CollectionResponse {
   name: string;
   path: string;
@@ -987,6 +992,62 @@ export async function handleUpdateCollection(
   });
 }
 
+/**
+ * POST /api/collections/:name/embeddings/clear
+ * Clear stale or all embeddings for a collection.
+ */
+export async function handleClearCollectionEmbeddings(
+  ctxHolder: ContextHolder,
+  store: SqliteAdapter,
+  name: string,
+  req: Request
+): Promise<Response> {
+  let body: ClearCollectionEmbeddingsRequestBody;
+  try {
+    body = (await req.json()) as ClearCollectionEmbeddingsRequestBody;
+  } catch {
+    return errorResponse("VALIDATION", "Invalid JSON body");
+  }
+
+  const mode = body.mode ?? "stale";
+  if (mode !== "stale" && mode !== "all") {
+    return errorResponse("VALIDATION", "mode must be 'stale' or 'all'");
+  }
+
+  const collection = ctxHolder.config.collections.find(
+    (item) => item.name === name.toLowerCase()
+  );
+  if (!collection) {
+    return errorResponse("NOT_FOUND", `Collection not found: ${name}`, 404);
+  }
+
+  const activeModel = resolveModelUri(
+    ctxHolder.config,
+    "embed",
+    undefined,
+    collection.name
+  );
+  const result = await store.clearEmbeddingsForCollection(collection.name, {
+    mode,
+    activeModel,
+  });
+  if (!result.ok) {
+    const status = result.error.code === "INVALID_INPUT" ? 400 : 500;
+    return errorResponse(result.error.code, result.error.message, status);
+  }
+
+  return jsonResponse({
+    success: true,
+    stats: result.value,
+    note:
+      mode === "all"
+        ? `Run gno embed --collection ${collection.name} to rebuild active embeddings.`
+        : result.value.protectedSharedVectors > 0
+          ? "Some shared vectors were retained because other active collections still use the same content."
+          : undefined,
+  });
+}
+
 /**
  * POST /api/sync
  * Trigger re-index of all or specific collection.
@@ -3356,6 +3417,7 @@ export async function handleSetPreset(
   }
 
   console.log(`Switching to preset: ${preset.name}`);
+  const previousEmbedModel = resolveModelUri(ctxHolder.config, "embed");
 
   const syncResult = await (deps?.applyConfigChangeFn ?? applyConfigChange)(
     ctxHolder,
@@ -3405,6 +3467,11 @@ export async function handleSetPreset(
   return jsonResponse({
     success: true,
     activePreset: body.presetId,
+    embedModelChanged: previousEmbedModel !== preset.embed,
+    note:
+      previousEmbedModel !== preset.embed
+        ? "Embedding model changed. Existing collections may need gno embed so vector results catch up."
+        : undefined,
     capabilities: ctxHolder.current.capabilities,
   });
 }

package/src/serve/server.ts

@@ -17,6 +17,7 @@ import {
   handleAsk,
   handleBrowseTree,
   handleCapabilities,
+  handleClearCollectionEmbeddings,
   handleCollections,
   handleConnectors,
   handleCreateFolder,
@@ -547,6 +548,25 @@ export async function startServer(
             );
           },
         },
+        "/api/collections/:name/embeddings/clear": {
+          POST: async (req: Request) => {
+            if (!isRequestAllowed(req, port)) {
+              return withSecurityHeaders(forbiddenResponse(), isDev);
+            }
+            const url = new URL(req.url);
+            const parts = url.pathname.split("/");
+            const name = decodeURIComponent(parts.at(-3) || "");
+            return withSecurityHeaders(
+              await handleClearCollectionEmbeddings(
+                ctxHolder,
+                store,
+                name,
+                req
+              ),
+              isDev
+            );
+          },
+        },
         "/api/doc/:id/links": {
           GET: async (req: Request) => {
             const url = new URL(req.url);

package/src/serve/status.ts

@@ -15,7 +15,7 @@ import type {
 import { getModelsCachePath } from "../app/constants";
 import { ModelCache } from "../llm/cache";
 import { envIsSet, resolveDownloadPolicy } from "../llm/policy";
-import { getActivePreset } from "../llm/registry";
+import { getActivePreset, resolveModelUri } from "../llm/registry";
 import { downloadState, type ServerContext } from "./context";
 
 const GIGABYTE = 1024 * 1024 * 1024;
@@ -615,7 +615,9 @@ export async function buildAppStatus(
   ctx: ServerContext,
   deps: StatusBuildDeps = {}
 ): Promise<AppStatusResponse> {
-  const result = await ctx.store.getStatus();
+  const result = await ctx.store.getStatus({
+    embedModel: resolveModelUri(ctx.config, "embed"),
+  });
   if (!result.ok) {
     throw result.error;
   }

package/src/store/sqlite/adapter.ts

@@ -24,6 +24,7 @@ import type {
   DocLinkSource,
   DocumentInput,
   DocumentRow,
+  EmbeddingCleanupStats,
   FtsResult,
   FtsSearchOptions,
   GetGraphOptions,
@@ -45,6 +46,7 @@ import { buildUri, deriveDocid } from "../../app/constants";
 import { normalizeWikiName, stripWikiMdExt } from "../../core/links";
 import { migrations, runMigrations } from "../migrations";
 import { err, ok } from "../types";
+import { modelTableName } from "../vector/sqlite-vec";
 import { loadFts5Snowball } from "./fts5-snowball";
 
 // ─────────────────────────────────────────────────────────────────────────────
@@ -2696,9 +2698,12 @@ export class SqliteAdapter implements StorePort, SqliteDbProvider {
   // Status
   // ─────────────────────────────────────────────────────────────────────────
 
-  async getStatus(): Promise<StoreResult<IndexStatus>> {
+  async getStatus(options?: {
+    embedModel?: string;
+  }): Promise<StoreResult<IndexStatus>> {
     try {
       const db = this.ensureOpen();
+      const embedModel = options?.embedModel ?? null;
 
       // Get version
       const versionRow = db
@@ -2729,7 +2734,7 @@ export class SqliteAdapter implements StorePort, SqliteDbProvider {
       }
 
       const collectionStats = db
-        .query<CollectionStat, []>(
+        .query<CollectionStat, [string | null, string | null]>(
           `
           SELECT
             c.name,
@@ -2741,15 +2746,26 @@ export class SqliteAdapter implements StorePort, SqliteDbProvider {
             (SELECT COUNT(*) FROM content_chunks cc
              JOIN documents d2 ON d2.mirror_hash = cc.mirror_hash
              WHERE d2.collection = c.name AND d2.active = 1) as chunk_count,
-            (SELECT COUNT(*) FROM content_vectors cv
-             JOIN documents d3 ON d3.mirror_hash = cv.mirror_hash
-             WHERE d3.collection = c.name AND d3.active = 1) as embedded_count
+            (SELECT COUNT(*) FROM content_chunks cc
+             WHERE EXISTS (
+               SELECT 1 FROM documents d3
+               WHERE d3.collection = c.name
+                 AND d3.active = 1
+                 AND d3.mirror_hash = cc.mirror_hash
+             )
+             AND EXISTS (
+               SELECT 1 FROM content_vectors cv
+               WHERE cv.mirror_hash = cc.mirror_hash
+                 AND cv.seq = cc.seq
+                 AND (? IS NULL OR cv.model = ?)
+                 AND cv.embedded_at >= cc.created_at
+             )) as embedded_count
           FROM collections c
           LEFT JOIN documents d ON d.collection = c.name
           GROUP BY c.name, c.path
         `
         )
-        .all();
+        .all(embedModel, embedModel);
 
       // Get totals
       const totalsRow = db
@@ -2773,7 +2789,7 @@ export class SqliteAdapter implements StorePort, SqliteDbProvider {
       // Embedding backlog: chunks from active docs without vectors
       // Uses EXISTS to avoid duplicates when multiple docs share mirror_hash
       const backlogRow = db
-        .query<{ count: number }, []>(
+        .query<{ count: number }, [string | null, string | null]>(
           `
           SELECT COUNT(*) as count FROM content_chunks c
           WHERE EXISTS (
@@ -2782,11 +2798,14 @@ export class SqliteAdapter implements StorePort, SqliteDbProvider {
           )
           AND NOT EXISTS (
             SELECT 1 FROM content_vectors v
-            WHERE v.mirror_hash = c.mirror_hash AND v.seq = c.seq
+            WHERE v.mirror_hash = c.mirror_hash
+              AND v.seq = c.seq
+              AND (? IS NULL OR v.model = ?)
+              AND v.embedded_at >= c.created_at
           )
         `
         )
-        .get();
+        .get(embedModel, embedModel);
 
       // Recent errors (last 24h)
       const recentErrorsRow = db
@@ -2960,6 +2979,126 @@ export class SqliteAdapter implements StorePort, SqliteDbProvider {
       );
     }
   }
+
+  async clearEmbeddingsForCollection(
+    collection: string,
+    options: { mode: "stale" | "all"; activeModel?: string }
+  ): Promise<StoreResult<EmbeddingCleanupStats>> {
+    try {
+      const db = this.ensureOpen();
+      const collectionName = collection.toLowerCase();
+
+      if (options.mode === "stale" && !options.activeModel) {
+        return err(
+          "INVALID_INPUT",
+          "activeModel is required for stale embedding cleanup"
+        );
+      }
+
+      const filterSql = options.mode === "stale" ? "AND cv.model != ?" : "";
+      const filterParams =
+        options.mode === "stale" ? [options.activeModel ?? ""] : [];
+
+      const deletableRows = db
+        .query<{ mirror_hash: string; model: string; seq: number }, string[]>(
+          `
+          SELECT DISTINCT cv.mirror_hash, cv.seq, cv.model
+          FROM content_vectors cv
+          WHERE EXISTS (
+            SELECT 1 FROM documents d
+            WHERE d.collection = ?
+              AND d.mirror_hash = cv.mirror_hash
+          )
+          ${filterSql}
+          AND NOT EXISTS (
+            SELECT 1 FROM documents d2
+            WHERE d2.mirror_hash = cv.mirror_hash
+              AND d2.collection != ?
+              AND d2.active = 1
+          )
+        `
+        )
+        .all(collectionName, ...filterParams, collectionName);
+
+      const protectedRow = db
+        .query<{ count: number }, string[]>(
+          `
+          SELECT COUNT(*) as count
+          FROM (
+            SELECT DISTINCT cv.mirror_hash, cv.seq, cv.model
+            FROM content_vectors cv
+            WHERE EXISTS (
+              SELECT 1 FROM documents d
+              WHERE d.collection = ?
+                AND d.mirror_hash = cv.mirror_hash
+            )
+            ${filterSql}
+            AND EXISTS (
+              SELECT 1 FROM documents d2
+              WHERE d2.mirror_hash = cv.mirror_hash
+                AND d2.collection != ?
+                AND d2.active = 1
+            )
+          )
+        `
+        )
+        .get(collectionName, ...filterParams, collectionName);
+
+      const deletedModels = [...new Set(deletableRows.map((row) => row.model))];
+
+      const transaction = db.transaction(() => {
+        const deleteVectorStmt = db.prepare(
+          `DELETE FROM content_vectors WHERE mirror_hash = ? AND seq = ? AND model = ?`
+        );
+        const vecDeleteStatements = new Map<
+          string,
+          ReturnType<typeof db.prepare> | null
+        >();
+
+        for (const row of deletableRows) {
+          deleteVectorStmt.run(row.mirror_hash, row.seq, row.model);
+
+          let deleteVecStmt = vecDeleteStatements.get(row.model);
+          if (deleteVecStmt === undefined) {
+            try {
+              deleteVecStmt = db.prepare(
+                `DELETE FROM ${modelTableName(row.model)} WHERE chunk_id = ?`
+              );
+            } catch {
+              deleteVecStmt = null;
+            }
+            vecDeleteStatements.set(row.model, deleteVecStmt);
+          }
+
+          if (deleteVecStmt) {
+            try {
+              deleteVecStmt.run(`${row.mirror_hash}:${row.seq}`);
+            } catch {
+              // Best effort; a later vec sync/rebuild can recover.
+            }
+          }
+        }
+      });
+
+      transaction();
+
+      return ok({
+        collection: collectionName,
+        deletedVectors: deletableRows.length,
+        deletedModels,
+        mode: options.mode,
+        protectedSharedVectors: protectedRow?.count ?? 0,
+      });
+    } catch (cause) {
+      return err(
+        "QUERY_FAILED",
+        cause instanceof Error
+          ? cause.message
+          : "Failed to clear collection embeddings",
+        cause
+      );
+    }
+  }
 }
 
 // ─────────────────────────────────────────────────────────────────────────────

package/src/store/types.ts

@@ -403,6 +403,14 @@ export interface CleanupStats {
   expiredCache: number;
 }
 
+export interface EmbeddingCleanupStats {
+  collection: string;
+  deletedVectors: number;
+  deletedModels: string[];
+  mode: "stale" | "all";
+  protectedSharedVectors: number;
+}
+
 // ─────────────────────────────────────────────────────────────────────────────
 // Graph Types
 // ─────────────────────────────────────────────────────────────────────────────
@@ -846,7 +854,9 @@ export interface StorePort {
   /**
    * Get index status with counts and health info.
    */
-  getStatus(): Promise<StoreResult<IndexStatus>>;
+  getStatus(options?: {
+    embedModel?: string;
+  }): Promise<StoreResult<IndexStatus>>;
 
   // ─────────────────────────────────────────────────────────────────────────
   // Errors
@@ -870,4 +880,12 @@ export interface StorePort {
    * Remove orphaned content, chunks, vectors, and expired cache.
    */
   cleanupOrphans(): Promise<StoreResult<CleanupStats>>;
+
+  /**
+   * Remove embeddings for a collection.
+   */
+  clearEmbeddingsForCollection(
+    collection: string,
+    options: { mode: "stale" | "all"; activeModel?: string }
+  ): Promise<StoreResult<EmbeddingCleanupStats>>;
 }

package/src/store/vector/sqlite-vec.ts

@@ -51,7 +51,7 @@ export function decodeEmbedding(blob: Uint8Array): Float32Array {
  * Generate deterministic table name from model URI.
  * First 8 chars of SHA256 hash.
  */
-function modelTableName(modelUri: string): string {
+export function modelTableName(modelUri: string): string {
   const hash = createHash("sha256").update(modelUri).digest("hex").slice(0, 8);
   return `vec_${hash}`;
 }