@gmickel/gno 0.37.0 → 0.39.0

package/README.md

@@ -85,45 +85,15 @@ gno daemon
 
 ---
 
-## What's New in v0.29
+## What's New
 
-- **GNO Desktop Beta**: first mac-first desktop beta shell with deep-link routing, singleton handoff, and the same onboarding/search/edit flows as `gno serve`
-- **Desktop Onboarding Polish**: guided setup now covers folders, presets, model readiness, indexing, connectors, import preview, app tabs, file actions, and recovery without drift between web and desktop
-- **Default Preset Upgrade**: `slim-tuned` is now the built-in default, using the fine-tuned retrieval expansion model while keeping the same embed, rerank, and answer stack as `slim`
-- **Workspace UI Polish**: richer scholarly-dusk presentation across dashboard, tabs, search, ask, footer, and global styling without introducing external font or asset dependencies
+> Latest release: [v0.37.0](./CHANGELOG.md#0370---2026-04-06)  
+> Full release history: [CHANGELOG.md](./CHANGELOG.md)
 
-## What's New in v0.30
-
-- **Headless Daemon Mode**: `gno daemon` keeps your index fresh continuously without opening the Web UI
-- **CLI Concurrency Hardening**: read-only commands no longer trip transient `database is locked` errors when they overlap with `gno update`
-- **Web/Desktop UI Polish**: sharper workspace styling across dashboard, tabs, search, ask, and footer surfaces
-
-## What's New in v0.31
-
-- **Windows Desktop Beta Artifact**: release flow now includes a packaged `windows-x64` desktop beta zip, not just source-level support claims
-- **Packaged Runtime Proof**: Windows desktop packaging validates bundled Bun + staged GNO runtime + FTS5 + vendored snowball + `sqlite-vec`
-- **Scoped Index Fix**: `gno index <collection>` now embeds only that collection instead of accidentally burning through unrelated backlog from other collections
-- **CLI Reporting Fix**: long embed runs now report sane durations instead of bogus sub-second summaries
-
-### v0.24
-
-- **Structured Query Documents**: first-class multi-line query syntax using `term:`, `intent:`, and `hyde:`
-- **Cross-Surface Rollout**: works across CLI, API, MCP, SDK, and Web Search/Ask
-- **Portable Retrieval Prompts**: save/share advanced retrieval intent as one text payload instead of repeated flags or JSON arrays
-
-### v0.23
-
-- **SDK / Library Mode**: package-root importable SDK with `createGnoClient(...)` for direct retrieval, document access, and indexing flows
-- **Inline Config Support**: embed GNO in another app without writing YAML config files
-- **Programmatic Indexing**: call `update`, `embed`, and `index` directly from Bun/TypeScript
-- **Docs & Website**: dedicated SDK guide, feature page, homepage section, and architecture docs
-
-### v0.22
-
-- **Promoted Slim Retrieval Model**: published `slim-retrieval-v1` on Hugging Face for direct `hf:` installation in GNO
-- **Fine-Tuning Workflow**: local MLX LoRA training, portable GGUF export, automatic checkpoint selection, promotion bundles, and repeatable benchmark comparisons
-- **Autonomous Search Harness**: bounded candidate search with early-stop guards, repeated incumbent confirmation, and promotion targets
-- **Public Docs & Site**: fine-tuned model docs and feature pages now point at the published HF model and the `slim-tuned` preset
+- **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
+- **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
 
@@ -133,7 +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
@@ -150,58 +120,6 @@ gno query "ECONNREFUSED 127.0.0.1:5432" --thorough
 
 > Full guide: [Fine-Tuned Models](https://gno.sh/docs/FINE-TUNED-MODELS/) · [Feature page](https://gno.sh/features/fine-tuned-models/)
 
-## What's New in v0.21
-
-- **Ask CLI Query Modes**: `gno ask` now accepts repeatable `--query-mode term|intent|hyde` entries, matching the existing Ask API and Web controls
-
-### v0.20
-
-- **Improved Model Init Fallbacks**: upgraded `node-llama-cpp` to `3.17.1` and switched to `build: "autoAttempt"` for better backend selection/fallback behavior
-
-### v0.19
-
-- **Exclusion Filters**: explicit `exclude` controls across CLI, API, Web, and MCP to hard-prune unwanted docs by title/path/body text
-- **Ask Query-Mode Parity**: Ask now supports structured `term` / `intent` / `hyde` controls in both API and Web UI
-
-### v0.18
-
-- **Intent Steering**: optional `intent` control for ambiguous queries across CLI, API, Web, and MCP query flows
-- **Rerank Controls**: `candidateLimit` lets you tune rerank cost vs. recall on slower or memory-constrained machines
-- **Stability**: query expansion now uses a bounded configurable context size (`models.expandContextSize`, default `2048`)
-- **Rerank Efficiency**: identical chunk texts are deduplicated before scoring and expanded back out deterministically
-
-### v0.17
-
-- **Structured Query Modes**: `term`, `intent`, and `hyde` controls across CLI, API, MCP, and Web
-- **Temporal Retrieval Upgrades**: `since`/`until`, date-range parsing, and recency sorting with frontmatter-date fallback
-- **Web Retrieval UX Polish**: richer advanced controls in Search and Ask (collection/date/category/author/tags + query modes)
-- **Metadata-Aware Retrieval**: ingestion now materializes document metadata/date fields for better filtering and ranking
-- **Migration Reliability**: SQLite-compatible migration path for existing indexes (including older SQLite engines)
-
-### v0.15
-
-- **HTTP Backends**: Offload embedding, reranking, and generation to remote GPU servers
-- Simple URI config: `http://host:port/path#modelname`
-- Works with llama-server, Ollama, LocalAI, vLLM
-- Run GNO on lightweight machines while GPU inference runs on your network
-
-### v0.13
-
-- **Knowledge Graph**: Interactive force-directed visualization of document connections
-- **Graph with Similarity**: See semantic similarity as golden edges (not just wiki/markdown links)
-- **CLI**: `gno graph` command with collection filtering and similarity options
-- **Web UI**: `/graph` page with zoom, pan, collection filter, similarity toggle
-- **MCP**: `gno_graph` tool for AI agents to explore document relationships
-- **REST API**: `/api/graph` endpoint with full query parameters
-
-### v0.12
-
-- **Note Linking**: Wiki-style `[[links]]`, backlinks, and AI-powered related notes
-- **Tag System**: Filter searches by frontmatter tags with `--tags-any`/`--tags-all`
-- **Web UI**: Outgoing links panel, backlinks panel, related notes sidebar
-- **CLI**: `gno links`, `gno backlinks`, `gno similar` commands
-- **MCP**: `gno_links`, `gno_backlinks`, `gno_similar` tools
-
 ---
 
 ## Quick Start
@@ -741,11 +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
 
@@ -896,7 +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

@@ -177,6 +177,35 @@ gno embed              # Embed only (if already synced)
 
 MCP `gno.sync` and `gno.capture` do NOT auto-embed. Use CLI for embedding.
 
+## Collection-specific embedding models
+
+Collections can override the global embedding model with `models.embed`.
+
+CLI path:
+
+```bash
+gno collection add ~/work/gno/src \
+  --name gno-code \
+  --embed-model "hf:Qwen/Qwen3-Embedding-0.6B-GGUF/Qwen3-Embedding-0.6B-Q8_0.gguf"
+```
+
+Good default guidance:
+
+- keep the global preset for mixed notes/docs collections
+- use a collection-specific embed override for code-heavy collections when benchmark guidance says so
+- after changing an embed model on an existing populated collection, run:
+
+```bash
+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

@@ -42,7 +42,11 @@ gno init [<path>] [options]
 gno collection add <path> --name <name> [options]
 ```
 
-Options same as `init`.
+Options same as `init`, plus:
+
+| Option                | Description                                          |
+| --------------------- | ---------------------------------------------------- |
+| `--embed-model <uri>` | Initial collection-specific embedding model override |
 
 ### gno collection list
 
@@ -62,6 +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": "0.37.0",
+  "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/add.ts

@@ -12,6 +12,7 @@ import {
 import { CliError } from "../../errors";
 
 interface AddOptions {
+  embedModel?: string;
   name?: string;
   pattern?: string;
   include?: string;
@@ -51,6 +52,11 @@ export async function collectionAdd(
     pattern: options.pattern,
     include: options.include,
     exclude: options.exclude,
+    models: options.embedModel
+      ? {
+          embed: options.embedModel,
+        }
+      : undefined,
     updateCmd: options.update,
   });
 

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

@@ -1264,10 +1264,12 @@ function wireManagementCommands(program: Command): void {
     .option("--pattern <glob>", "file matching pattern")
     .option("--include <exts>", "extension allowlist (CSV)")
     .option("--exclude <patterns>", "exclude patterns (CSV)")
+    .option("--embed-model <uri>", "collection-specific embedding model URI")
     .option("--update <cmd>", "shell command to run before indexing")
     .action(async (path: string, cmdOpts: Record<string, unknown>) => {
       const { collectionAdd } = await import("./commands/collection");
       await collectionAdd(path, {
+        embedModel: cmdOpts.embedModel as string | undefined,
         name: cmdOpts.name as string,
         pattern: cmdOpts.pattern as string | undefined,
         include: cmdOpts.include as string | undefined,
@@ -1308,6 +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/collection/add.ts

@@ -97,6 +97,7 @@ export async function addCollection(
     pattern: input.pattern ?? DEFAULT_PATTERN,
     include: includeList,
     exclude: excludeList,
+    models: input.models,
     updateCmd: input.updateCmd,
   };
 

package/src/collection/index.ts

@@ -7,6 +7,7 @@
 
 export { addCollection } from "./add";
 export { removeCollection } from "./remove";
+export { updateCollection } from "./update";
 export type {
   AddCollectionInput,
   CollectionError,
@@ -14,4 +15,5 @@ export type {
   CollectionSuccess,
   RemoveCollectionInput,
   RenameCollectionInput,
+  UpdateCollectionInput,
 } from "./types";

package/src/collection/types.ts

@@ -5,6 +5,7 @@
  */
 
 import type { Collection, Config } from "../config/types";
+import type { ModelType } from "../llm/types";
 
 /**
  * Input for adding a collection.
@@ -22,6 +23,8 @@ export interface AddCollectionInput {
   exclude?: string[] | string;
   /** Update command to run before sync */
   updateCmd?: string;
+  /** Optional initial model overrides */
+  models?: Partial<Record<ModelType, string>>;
 }
 
 /**
@@ -42,6 +45,16 @@ export interface RenameCollectionInput {
   newName: string;
 }
 
+/**
+ * Input for updating a collection.
+ */
+export interface UpdateCollectionInput {
+  /** Collection name (case-insensitive) */
+  name: string;
+  /** Partial model override patch; null clears one role */
+  models?: Partial<Record<ModelType, string | null>>;
+}
+
 /**
  * Successful collection operation result.
  */

package/src/collection/update.ts

@@ -0,0 +1,93 @@
+/**
+ * Update collection core logic.
+ * Pure function that mutates config - caller handles I/O.
+ *
+ * @module src/collection/update
+ */
+
+import type {
+  Collection,
+  CollectionModelOverrides,
+  Config,
+} from "../config/types";
+import type { CollectionResult, UpdateCollectionInput } from "./types";
+
+import { CollectionSchema } from "../config";
+
+function normalizeOverrides(
+  models?: UpdateCollectionInput["models"]
+): CollectionModelOverrides | undefined {
+  if (!models) {
+    return undefined;
+  }
+
+  const entries = Object.entries(models).filter(
+    ([, value]) => value !== undefined && value !== null
+  );
+  if (entries.length === 0) {
+    return undefined;
+  }
+
+  return Object.fromEntries(entries) as CollectionModelOverrides;
+}
+
+/**
+ * Update a collection in config.
+ */
+export function updateCollection(
+  config: Config,
+  input: UpdateCollectionInput
+): CollectionResult {
+  const collectionName = input.name.toLowerCase();
+  const index = config.collections.findIndex((c) => c.name === collectionName);
+  if (index < 0) {
+    return {
+      ok: false,
+      code: "NOT_FOUND",
+      message: `Collection "${collectionName}" not found`,
+    };
+  }
+
+  const current = config.collections[index];
+  if (!current) {
+    return {
+      ok: false,
+      code: "NOT_FOUND",
+      message: `Collection "${collectionName}" not found`,
+    };
+  }
+
+  let nextModels = current.models;
+  if (input.models) {
+    nextModels = normalizeOverrides({
+      ...current.models,
+      ...input.models,
+    });
+  }
+
+  const nextCollection: Collection = {
+    ...current,
+    models: nextModels,
+  };
+
+  const validation = CollectionSchema.safeParse(nextCollection);
+  if (!validation.success) {
+    return {
+      ok: false,
+      code: "VALIDATION",
+      message: `Invalid collection: ${validation.error.issues[0]?.message ?? "unknown error"}`,
+    };
+  }
+
+  const nextCollections = [...config.collections];
+  nextCollections[index] = validation.data;
+
+  return {
+    ok: true,
+    config: {
+      ...config,
+      collections: nextCollections,
+    },
+    collection: validation.data,
+  };
+}

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: