@gmickel/gno 0.36.0 → 0.38.0

package/README.md

@@ -1,6 +1,6 @@
 # GNO
 
-**Your Local Second Brain**: Index, search, and synthesize your entire digital life.
+**Local search, retrieval, and synthesis for the files you actually work in.**
 
 [![npm](./assets/badges/npm.svg)](https://www.npmjs.com/package/@gmickel/gno)
 [![MIT License](./assets/badges/license.svg)](./LICENSE)
@@ -12,7 +12,57 @@
 
 ![GNO](./assets/og-image.png)
 
-GNO is a local knowledge engine that turns your documents into a searchable, connected knowledge graph. Index notes, code, PDFs, and Office docs. Get hybrid search, AI answers with citations, and wiki-style note linking—all 100% offline.
+GNO is a local knowledge engine for notes, code, PDFs, Office docs, meeting transcripts, and reference material. It gives you fast keyword search, semantic retrieval, grounded answers with citations, wiki-style linking, and a real workspace UI, while keeping the whole stack local by default.
+
+Use it when:
+
+- your notes live in more than one folder
+- your important knowledge is split across Markdown, code, PDFs, and Office files
+- you want one retrieval layer that works from the CLI, browser, MCP, and a Bun/TypeScript SDK
+- you want better local context for agents without shipping your docs to a cloud API
+
+### What GNO Gives You
+
+- **Fast local search**: BM25 for exact hits, vectors for concepts, hybrid for best quality
+- **Real retrieval surfaces**: CLI, Web UI, REST API, MCP, SDK
+- **Local-first answers**: grounded synthesis with citations when you want answers, raw retrieval when you do not
+- **Connected knowledge**: backlinks, related notes, graph view, cross-collection navigation
+- **Operational fit**: daemon mode, model presets, remote GPU backends, safe config/state on disk
+
+### One-Minute Tour
+
+```bash
+# Install
+bun install -g @gmickel/gno
+
+# Add a few collections
+gno init ~/notes --name notes
+gno collection add ~/work/docs --name work-docs --pattern "**/*.{md,pdf,docx}"
+gno collection add ~/work/gno/src --name gno-code --pattern "**/*.{ts,tsx,js,jsx}"
+
+# Add context so retrieval results come back with the right framing
+gno context add "notes:" "Personal notes, journal entries, and long-form ideas"
+gno context add "work-docs:" "Architecture docs, runbooks, RFCs, meeting notes"
+gno context add "gno-code:" "Source code for the GNO application"
+
+# Index + embed
+gno update --yes
+gno embed
+
+# Search in the way that fits the question
+gno search "DEC-0054"                            # exact keyword / identifier
+gno vsearch "retry failed jobs with backoff"     # natural-language semantic lookup
+gno query "JWT refresh token rotation" --explain # hybrid retrieval with score traces
+
+# Retrieve documents or export context for an agent
+gno get "gno://work-docs/architecture/auth.md"
+gno multi-get "gno-code/**/*.ts" --max-bytes 30000 --md
+gno query "deployment process" --all --files --min-score 0.35
+
+# Run the workspace
+gno serve
+gno daemon
+```
 
 ---
 
@@ -35,45 +85,15 @@ GNO is a local knowledge engine that turns your documents into a searchable, con
 
 ---
 
-## What's New in v0.29
-
-- **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
-
-## 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
+## What's New
 
-### v0.24
+> Latest release: [v0.37.0](./CHANGELOG.md#0370---2026-04-06)  
+> Full release history: [CHANGELOG.md](./CHANGELOG.md)
 
-- **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
+- **Recommended Code Embed Model**: docs and benchmark pages now point to `Qwen3-Embedding-0.6B-GGUF` as the current code-specialist embedding option
+- **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
 
@@ -100,58 +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
@@ -265,6 +233,14 @@ headless. In v0.30 it is foreground-only and does not expose built-in
 
 Embed GNO directly in another Bun or TypeScript app. No CLI subprocesses. No local server required.
 
+Install:
+
+```bash
+bun add @gmickel/gno
+```
+
+Minimal client:
+
 ```ts
 import { createDefaultConfig, createGnoClient } from "@gmickel/gno";
 
@@ -295,6 +271,43 @@ console.log(results.results[0]?.uri);
 await client.close();
 ```
 
+More SDK examples:
+
+```ts
+import { createGnoClient } from "@gmickel/gno";
+
+const client = await createGnoClient({
+  configPath: "/Users/me/.config/gno/index.yml",
+});
+
+// Fast exact search
+const bm25 = await client.search("DEC-0054", {
+  collection: "work-docs",
+});
+
+// Semantic code lookup
+const semantic = await client.vsearch("retry failed jobs with backoff", {
+  collection: "gno-code",
+});
+
+// Hybrid retrieval with explicit intent
+const hybrid = await client.query("token refresh", {
+  collection: "work-docs",
+  intent: "JWT refresh token rotation in our auth stack",
+  candidateLimit: 12,
+});
+
+// Fetch content directly
+const doc = await client.get("gno://work-docs/auth/refresh.md");
+const bundle = await client.multiGet(["gno-code/**/*.ts"], { maxBytes: 25000 });
+
+// Indexing / embedding
+await client.update({ collection: "work-docs" });
+await client.embed({ collection: "gno-code" });
+
+await client.close();
+```
+
 Core SDK surface:
 
 - `createGnoClient({ config | configPath, dbPath? })`
@@ -303,12 +316,6 @@ Core SDK surface:
 - `update`, `embed`, `index`
 - `close`
 
-Install in an app:
-
-```bash
-bun add @gmickel/gno
-```
-
 Full guide: [SDK docs](https://gno.sh/docs/SDK/)
 
 ---
@@ -338,6 +345,31 @@ gno ask "what did we decide" --answer # AI synthesis
 
 Output formats: `--json`, `--files`, `--csv`, `--md`, `--xml`
 
+### Common CLI Recipes
+
+```bash
+# Search one collection
+gno search "PostgreSQL connection pool" --collection work-docs
+
+# Export retrieval results for an agent
+gno query "authentication flow" --json -n 10
+gno query "deployment rollback" --all --files --min-score 0.4
+
+# Retrieve a document by URI or docid
+gno get "gno://work-docs/runbooks/deploy.md"
+gno get "#abc123"
+
+# Fetch many documents at once
+gno multi-get "work-docs/**/*.md" --max-bytes 20000 --md
+
+# Inspect how the hybrid rank was assembled
+gno query "refresh token rotation" --explain
+
+# Work with filters
+gno query "meeting notes" --since "last month" --category "meeting,notes"
+gno search "incident review" --tags-all "status/active,team/platform"
+```
+
 ### Retrieval V2 Controls
 
 Existing query calls still work. Retrieval v2 adds optional structured intent control and deeper explain output.
@@ -382,6 +414,20 @@ gno skill install --scope user
 
 Then ask your agent: _"Search my notes for the auth discussion"_
 
+Agent-friendly CLI examples:
+
+```bash
+# Structured retrieval output for an agent
+gno query "authentication" --json -n 10
+
+# File list for downstream retrieval
+gno query "error handling" --all --files --min-score 0.35
+
+# Full document content when the agent already knows the ref
+gno get "gno://work-docs/api-reference.md" --full
+gno multi-get "work-docs/**/*.md" --md --max-bytes 30000
+```
+
 [Skill setup guide →](https://gno.sh/docs/integrations/skills/)
 
 ### MCP Server
@@ -655,7 +701,7 @@ See:
 Offload inference to a GPU server on your network:
 
 ```yaml
-# ~/.config/gno/config.yaml
+# ~/.config/gno/index.yml
 models:
   activePreset: remote-gpu
   presets:
@@ -715,6 +761,61 @@ bun run eval:hybrid:delta
 - Benchmark guide: [evals/README.md](./evals/README.md)
 - Latest baseline snapshot: [evals/fixtures/hybrid-baseline/latest.json](./evals/fixtures/hybrid-baseline/latest.json)
 
+### Code Embedding Benchmark Harness
+
+GNO also has a dedicated harness for comparing alternate embedding models on code retrieval without touching product defaults:
+
+```bash
+# Establish the current incumbent baseline
+bun run bench:code-embeddings --candidate bge-m3-incumbent --write
+
+# Add candidate model URIs to the search space, then inspect them
+bun run research:embeddings:autonomous:list-search-candidates
+
+# Benchmark one candidate explicitly
+bun run research:embeddings:autonomous:run-candidate bge-m3-incumbent
+
+# Or let the bounded search harness walk the remaining candidates later
+bun run research:embeddings:autonomous:search --dry-run
+```
+
+See [research/embeddings/README.md](./research/embeddings/README.md).
+
+If a model turns out to be better specifically for code, the intended user story is:
+
+- keep the default global preset for mixed prose/docs collections
+- use per-collection `models.embed` overrides for code collections
+
+That lets GNO stay sane by default while still giving power users a clean path to code-specialist retrieval.
+
+Current code-focused recommendation:
+
+```yaml
+collections:
+  - name: gno-code
+    path: /Users/you/work/gno/src
+    pattern: "**/*.{ts,tsx,js,jsx,go,rs,py,swift,c}"
+    models:
+      embed: "hf:Qwen/Qwen3-Embedding-0.6B-GGUF/Qwen3-Embedding-0.6B-Q8_0.gguf"
+```
+
+GNO treats that override like any other model URI:
+
+- auto-downloads on first use by default
+- manual-only if `GNO_NO_AUTO_DOWNLOAD=1`
+- offline-safe if the model is already cached
+
+Why this is the current recommendation:
+
+- matches `bge-m3` on the tiny canonical benchmark
+- significantly beats `bge-m3` on the real GNO `src/serve` code slice
+- also beats `bge-m3` on a pinned public-OSS code slice
+
+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
+
 ---
 
 ## License

package/assets/skill/SKILL.md

@@ -177,6 +177,28 @@ 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
+```
+
 ## 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
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gmickel/gno",
-  "version": "0.36.0",
+  "version": "0.38.0",
   "description": "Local semantic search for your documents. Index Markdown, PDF, and Office files with hybrid BM25 + vector search.",
   "keywords": [
     "embeddings",
@@ -69,6 +69,8 @@
     "eval:hybrid": "bun --bun evalite evals/hybrid.eval.ts",
     "eval:hybrid:baseline": "bun scripts/hybrid-benchmark.ts --write",
     "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",
     "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",
@@ -83,6 +85,11 @@
     "research:finetune:autonomous:confirm-winner": "bun research/finetune/autonomous/scripts/confirm-winner.ts",
     "research:finetune:autonomous:check-promotion-targets": "bun research/finetune/autonomous/scripts/check-promotion-targets.ts",
     "research:finetune:validate": "bun research/finetune/scripts/validate-sandbox.ts",
+    "research:embeddings:autonomous:list-search-candidates": "bun research/embeddings/autonomous/scripts/list-search-candidates.ts",
+    "research:embeddings:autonomous:run-candidate": "bun research/embeddings/autonomous/scripts/run-candidate.ts",
+    "research:embeddings:autonomous:leaderboard": "bun research/embeddings/autonomous/scripts/leaderboard.ts",
+    "research:embeddings:autonomous:confirm-winner": "bun research/embeddings/autonomous/scripts/confirm-winner.ts",
+    "research:embeddings:autonomous:search": "bun research/embeddings/autonomous/scripts/search.ts",
     "research:finetune:qmd-import:legacy": "bun research/finetune/scripts/import-qmd-training.ts",
     "research:finetune:mlx:build-dataset": "bun research/finetune/scripts/build-mlx-dataset.ts",
     "research:finetune:build-variant-dataset": "bun research/finetune/scripts/build-variant-dataset.ts",

package/src/cli/commands/ask.ts

@@ -14,7 +14,7 @@ import type { AskOptions, AskResult, Citation } from "../../pipeline/types";
 
 import { LlmAdapter } from "../../llm/nodeLlamaCpp/adapter";
 import { resolveDownloadPolicy } from "../../llm/policy";
-import { getActivePreset } from "../../llm/registry";
+import { resolveModelUri } from "../../llm/registry";
 import {
   generateGroundedAnswer,
   processAnswerResult,
@@ -90,7 +90,6 @@ export async function ask(
   let rerankPort: RerankPort | null = null;
 
   try {
-    const preset = getActivePreset(config);
     const llm = new LlmAdapter(config);
 
     // Resolve download policy from env/flags
@@ -106,7 +105,12 @@ export async function ask(
       : undefined;
 
     // Create embedding port
-    const embedUri = options.embedModel ?? preset.embed;
+    const embedUri = resolveModelUri(
+      config,
+      "embed",
+      options.embedModel,
+      options.collection
+    );
     const embedResult = await llm.createEmbeddingPort(embedUri, {
       policy,
       onProgress: downloadProgress
@@ -119,8 +123,12 @@ export async function ask(
 
     // Create expansion port when expansion is enabled.
     if (!options.noExpand && !options.queryModes?.length) {
-      const expandUri =
-        options.expandModel ?? options.genModel ?? preset.expand;
+      const expandUri = resolveModelUri(
+        config,
+        "expand",
+        options.expandModel ?? options.genModel,
+        options.collection
+      );
       const genResult = await llm.createExpansionPort(expandUri, {
         policy,
         onProgress: downloadProgress
@@ -134,7 +142,12 @@ export async function ask(
 
     // Create answer generation port when answers are requested.
     if (options.answer) {
-      const genUri = options.genModel ?? preset.gen;
+      const genUri = resolveModelUri(
+        config,
+        "gen",
+        options.genModel,
+        options.collection
+      );
       const genResult = await llm.createGenerationPort(genUri, {
         policy,
         onProgress: downloadProgress
@@ -148,7 +161,12 @@ export async function ask(
 
     // Create rerank port (unless --fast or --no-rerank)
     if (!options.noRerank) {
-      const rerankUri = options.rerankModel ?? preset.rerank;
+      const rerankUri = resolveModelUri(
+        config,
+        "rerank",
+        options.rerankModel,
+        options.collection
+      );
       const rerankResult = await llm.createRerankPort(rerankUri, {
         policy,
         onProgress: downloadProgress

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/doctor.ts

@@ -14,6 +14,7 @@ import type { Config } from "../../config/types";
 
 import { getIndexDbPath, getModelsCachePath } from "../../app/constants";
 import { getConfigPaths, isInitialized, loadConfig } from "../../config";
+import { getCodeChunkingStatus } from "../../ingestion/chunker";
 import { ModelCache } from "../../llm/cache";
 import { getActivePreset } from "../../llm/registry";
 import { loadFts5Snowball } from "../../store/sqlite/fts5-snowball";
@@ -122,6 +123,19 @@ async function checkModels(config: Config): Promise<DoctorCheck[]> {
   return checks;
 }
 
+function checkCodeChunking(): DoctorCheck {
+  const status = getCodeChunkingStatus();
+  return {
+    name: "code-chunking",
+    status: "ok",
+    message: `${status.mode} structural chunking for ${status.supportedExtensions.join(", ")}`,
+    details: [
+      "Unsupported extensions fall back to the default markdown chunker.",
+      "Chunking mode is automatic-only in the first pass.",
+    ],
+  };
+}
+
 async function checkNodeLlamaCpp(): Promise<DoctorCheck> {
   try {
     const { getLlama } = await import("node-llama-cpp");
@@ -319,6 +333,9 @@ export async function doctor(
   const sqliteChecks = await checkSqliteExtensions();
   checks.push(...sqliteChecks);
 
+  // Code chunking capability
+  checks.push(checkCodeChunking());
+
   // Determine overall health
   const hasErrors = checks.some((c) => c.status === "error");
 

package/src/cli/commands/embed.ts

@@ -19,7 +19,7 @@ import {
 } from "../../config";
 import { LlmAdapter } from "../../llm/nodeLlamaCpp/adapter";
 import { resolveDownloadPolicy } from "../../llm/policy";
-import { getActivePreset } from "../../llm/registry";
+import { resolveModelUri } from "../../llm/registry";
 import { formatDocForEmbedding } from "../../pipeline/contextual";
 import { SqliteAdapter } from "../../store/sqlite/adapter";
 import { err, ok } from "../../store/types";
@@ -271,8 +271,7 @@ async function initEmbedContext(
     return { ok: false, error: `Collection not found: ${collection}` };
   }
 
-  const preset = getActivePreset(config);
-  const modelUri = model ?? preset.embed;
+  const modelUri = resolveModelUri(config, "embed", model, collection);
 
   const store = new SqliteAdapter();
   const dbPath = getIndexDbPath();

package/src/cli/commands/query.ts

@@ -14,7 +14,7 @@ import type { HybridSearchOptions, SearchResults } from "../../pipeline/types";
 
 import { LlmAdapter } from "../../llm/nodeLlamaCpp/adapter";
 import { resolveDownloadPolicy } from "../../llm/policy";
-import { getActivePreset } from "../../llm/registry";
+import { resolveModelUri } from "../../llm/registry";
 import { type HybridSearchDeps, searchHybrid } from "../../pipeline/hybrid";
 import {
   createVectorIndexPort,
@@ -58,6 +58,7 @@ export interface QueryFormatOptions {
   format: "terminal" | "json" | "files" | "csv" | "md" | "xml";
   full?: boolean;
   lineNumbers?: boolean;
+  terminalLinks?: import("../format/search-results").FormatOptions["terminalLinks"];
 }
 
 export type QueryResult =
@@ -97,7 +98,6 @@ export async function query(
   let rerankPort: RerankPort | null = null;
 
   try {
-    const preset = getActivePreset(config);
     const llm = new LlmAdapter(config);
 
     // Resolve download policy from env/flags
@@ -113,7 +113,12 @@ export async function query(
       : undefined;
 
     // Create embedding port (for vector search)
-    const embedUri = options.embedModel ?? preset.embed;
+    const embedUri = resolveModelUri(
+      config,
+      "embed",
+      options.embedModel,
+      options.collection
+    );
     const embedResult = await llm.createEmbeddingPort(embedUri, {
       policy,
       onProgress: downloadProgress
@@ -127,8 +132,12 @@ export async function query(
     // Create expansion port - optional.
     // Skip when structured query modes are provided.
     if (!options.noExpand && !options.queryModes?.length) {
-      const expandUri =
-        options.expandModel ?? options.genModel ?? preset.expand;
+      const expandUri = resolveModelUri(
+        config,
+        "expand",
+        options.expandModel ?? options.genModel,
+        options.collection
+      );
       const genResult = await llm.createExpansionPort(expandUri, {
         policy,
         onProgress: downloadProgress
@@ -142,7 +151,12 @@ export async function query(
 
     // Create rerank port - optional
     if (!options.noRerank) {
-      const rerankUri = options.rerankModel ?? preset.rerank;
+      const rerankUri = resolveModelUri(
+        config,
+        "rerank",
+        options.rerankModel,
+        options.collection
+      );
       const rerankResult = await llm.createRerankPort(rerankUri, {
         policy,
         onProgress: downloadProgress
@@ -260,5 +274,6 @@ export function formatQuery(
     format: options.format,
     full: options.full,
     lineNumbers: options.lineNumbers,
+    terminalLinks: options.terminalLinks,
   });
 }