codesearch-mcp 0.1.0-alpha → 0.2.1-alpha

package/dist/src/tools/GetCallersTool.js

@@ -10,11 +10,15 @@ class GetCallersTool {
     definition() {
         return {
             name: 'get_callers',
-            description: 'Find all call sites of a function or method by exact name. Uses the call graph index — faster and more precise than grep. Use after find_symbol to trace where a symbol is actually called. Returns empty if rag-index backfill --calls has not been run.',
+            description: 'Find all call sites of a function or method by exact name. Accepts a bare name (every call site with that ' +
+                'callee name) or a qualified name (`Class.method`, from find_symbol output) to disambiguate same-named methods ' +
+                'of different classes. Uses the call graph index — faster and more precise than grep. A call through an ' +
+                'interface-typed receiver is recorded under the interface member (e.g. `IEmbedder.embedBatch`), not the ' +
+                'implementing class. Use after find_symbol to trace where a symbol is actually called.',
             inputSchema: {
                 type: 'object',
                 properties: {
-                    symbol: { type: 'string', description: 'Exact function or method name (case-sensitive)' },
+                    symbol: { type: 'string', description: 'Bare or qualified (Class.method) function/method name (case-sensitive)' },
                     project: { type: 'string', description: 'Project name (from list_projects)' },
                     callee_file: { type: 'string', description: 'Optional: filter by definition file path (as returned by find_symbol) to disambiguate overloaded names' },
                 },
@@ -26,11 +30,22 @@ class GetCallersTool {
         const symbol = String(args.symbol ?? '');
         const project = String(args.project ?? '');
         const calleeFile = args.callee_file ? String(args.callee_file) : null;
-        const rows = await this.searcher.getCallers(symbol, project, calleeFile);
+        let rows;
+        try {
+            rows = await this.searcher.getCallers(symbol, project, calleeFile);
+        }
+        catch (err) {
+            return (0, format_helpers_1.errorResponse)(err.message);
+        }
+        const notice = await this.searcher.outdatedNotice(project);
+        const lines = [];
+        if (notice)
+            lines.push(notice + '\n');
         if (rows.length === 0) {
-            return (0, format_helpers_1.textResponse)(`No call sites found for \`${symbol}\` in project \`${project}\`. Run \`rag-index backfill --calls\` if the call graph has not been populated yet.`);
+            lines.push(`No call sites found for \`${symbol}\` in project \`${project}\`. Run \`rag-index --reindex\` if the call graph has not been populated yet.`);
+            return (0, format_helpers_1.textResponse)(lines.join('\n'));
         }
-        const lines = [`**Call sites of \`${symbol}\`** (${rows.length} found)\n`];
+        lines.push(`**Call sites of \`${symbol}\`** (${rows.length} found)\n`);
         for (const r of rows) {
             const loc = r.line != null ? ` line ${r.line}` : '';
             const ctx = r.fromSymbol ? `, in \`${r.fromSymbol}\`` : '';

package/docker/docker-compose.yml

@@ -12,19 +12,12 @@ services:
     volumes:
       - pgdata:/var/lib/postgresql/data
 
-  ollama:
-    image: ollama/ollama
-    ports:
-      - "11434:11434"
-    volumes:
-      - ollama_data:/root/.ollama
-
   # One Infinity container serves both /embeddings and /rerank. The model set is
   # bound at startup via INFINITY_MODEL_ID (";"-separated for multiple models,
-  # e.g. "BAAI/bge-m3;BAAI/bge-reranker-v2-m3"). Defaults to CPU so it starts on
-  # any host; GPU is opt-in via docker-compose.infinity-gpu.yml (see rag-setup).
-  # Behind an "infinity" profile so `docker compose up -d` (ollama path) doesn't
-  # pull this multi-GB image — it starts only when explicitly targeted.
+  # e.g. "BAAI/bge-m3;BAAI/bge-reranker-v2-m3") and rewritten in place by the app
+  # (a literal, not an env interpolation) so a plain `docker compose up -d`
+  # always serves the configured set. Defaults to CPU so it starts on any host;
+  # GPU is opt-in via docker-compose.infinity-gpu.yml (see rag-setup).
   # Model warmup defaults OFF here: on CPU it runs multi-minute benchmark
   # batches at every boot (a pure perf measurement, not required to serve). The
   # GPU override turns it back on, where it costs seconds.
@@ -34,16 +27,13 @@ services:
     ports:
       - "7997:7997"
     environment:
-      INFINITY_MODEL_ID: ${INFINITY_MODEL_ID:-jinaai/jina-embeddings-v2-base-code}
+      INFINITY_MODEL_ID: "jinaai/jina-embeddings-v2-base-code;mixedbread-ai/mxbai-rerank-xsmall-v1"
       INFINITY_MODEL_WARMUP: ${INFINITY_MODEL_WARMUP:-false}
       INFINITY_PORT: "7997"
     command: ["v2"]
     volumes:
       - infinity_cache:/app/.cache
-    profiles:
-      - infinity
 
 volumes:
   pgdata:
-  ollama_data:
   infinity_cache:

package/docs/architecture.svg

@@ -0,0 +1,103 @@
+<svg width="840" height="320" viewBox="0 0 840 320" xmlns="http://www.w3.org/2000/svg" font-family="system-ui,-apple-system,sans-serif">
+  <defs>
+    <marker id="a" markerWidth="8" markerHeight="6" refX="7" refY="3" orient="auto">
+      <polygon points="0 0,8 3,0 6" fill="#64748b"/>
+    </marker>
+    <marker id="ad" markerWidth="8" markerHeight="6" refX="7" refY="3" orient="auto">
+      <polygon points="0 0,8 3,0 6" fill="#94a3b8"/>
+    </marker>
+  </defs>
+
+  <!-- Background -->
+  <rect width="840" height="320" fill="#f8fafc" rx="12"/>
+
+  <!-- ── INDEXING lane ───────────────────────────────────────────── -->
+  <rect x="12" y="12" width="816" height="128" rx="8" fill="#eff6ff" stroke="#bfdbfe" stroke-width="1"/>
+  <text x="24" y="30" font-size="10" fill="#6b7280" font-weight="700" letter-spacing="1.2">INDEXING</text>
+
+  <!-- Source files -->
+  <rect x="30" y="42" width="104" height="52" rx="6" fill="#fff" stroke="#3b82f6" stroke-width="1.5"/>
+  <text x="82" y="64" font-size="11" text-anchor="middle" fill="#1e40af" font-weight="600">Source files</text>
+  <text x="82" y="80" font-size="10" text-anchor="middle" fill="#94a3b8">.ts · .tsx · .js</text>
+
+  <line x1="134" y1="68" x2="158" y2="68" stroke="#64748b" stroke-width="1.5" marker-end="url(#a)"/>
+
+  <!-- AST parser -->
+  <rect x="160" y="42" width="120" height="52" rx="6" fill="#fff" stroke="#3b82f6" stroke-width="1.5"/>
+  <text x="220" y="64" font-size="11" text-anchor="middle" fill="#1e40af" font-weight="600">AST parser</text>
+  <text x="220" y="80" font-size="10" text-anchor="middle" fill="#94a3b8">symbols · imports</text>
+
+  <line x1="280" y1="68" x2="304" y2="68" stroke="#64748b" stroke-width="1.5" marker-end="url(#a)"/>
+
+  <!-- Chunker -->
+  <rect x="306" y="42" width="110" height="52" rx="6" fill="#fff" stroke="#3b82f6" stroke-width="1.5"/>
+  <text x="361" y="64" font-size="11" text-anchor="middle" fill="#1e40af" font-weight="600">Chunker</text>
+  <text x="361" y="80" font-size="10" text-anchor="middle" fill="#94a3b8">symbol-aligned</text>
+
+  <line x1="416" y1="68" x2="440" y2="68" stroke="#64748b" stroke-width="1.5" marker-end="url(#a)"/>
+
+  <!-- Embedder -->
+  <rect x="442" y="42" width="128" height="52" rx="6" fill="#fff" stroke="#3b82f6" stroke-width="1.5"/>
+  <text x="506" y="61" font-size="11" text-anchor="middle" fill="#1e40af" font-weight="600">Embedder</text>
+  <text x="506" y="76" font-size="10" text-anchor="middle" fill="#94a3b8">Infinity · OpenAI</text>
+  <text x="506" y="89" font-size="10" text-anchor="middle" fill="#94a3b8">Voyage</text>
+
+  <line x1="570" y1="68" x2="610" y2="68" stroke="#64748b" stroke-width="1.5" marker-end="url(#a)"/>
+
+  <!-- PostgreSQL -->
+  <rect x="612" y="36" width="200" height="116" rx="8" fill="#faf5ff" stroke="#7c3aed" stroke-width="1.5"/>
+  <text x="712" y="58" font-size="11" text-anchor="middle" fill="#5b21b6" font-weight="700">PostgreSQL + pgvector</text>
+  <line x1="628" y1="66" x2="796" y2="66" stroke="#c4b5fd" stroke-width="1"/>
+  <text x="712" y="84" font-size="10" text-anchor="middle" fill="#7c3aed">code_chunk  (vectors)</text>
+  <text x="712" y="100" font-size="10" text-anchor="middle" fill="#7c3aed">code_symbol</text>
+  <text x="712" y="116" font-size="10" text-anchor="middle" fill="#7c3aed">code_dependency · code_call</text>
+
+  <!-- AST → DB (graph data) dashed -->
+  <path d="M220 94 Q220 136 612 136" stroke="#94a3b8" stroke-width="1" stroke-dasharray="4,3" fill="none" marker-end="url(#ad)"/>
+  <text x="400" y="150" font-size="10" text-anchor="middle" fill="#94a3b8">symbols · import graph · call graph</text>
+
+  <!-- ── QUERY lane ──────────────────────────────────────────────── -->
+  <rect x="12" y="182" width="816" height="126" rx="8" fill="#f0fdf4" stroke="#bbf7d0" stroke-width="1"/>
+  <text x="24" y="200" font-size="10" fill="#6b7280" font-weight="700" letter-spacing="1.2">QUERY</text>
+
+  <!-- Claude -->
+  <ellipse cx="72" cy="248" rx="52" ry="28" fill="#fff" stroke="#f97316" stroke-width="1.5"/>
+  <text x="72" y="244" font-size="11" text-anchor="middle" fill="#9a3412" font-weight="600">Claude</text>
+  <text x="72" y="259" font-size="10" text-anchor="middle" fill="#94a3b8">AI agent</text>
+
+  <line x1="124" y1="248" x2="152" y2="248" stroke="#64748b" stroke-width="1.5" marker-end="url(#a)"/>
+
+  <!-- rag-mcp -->
+  <rect x="154" y="224" width="110" height="48" rx="6" fill="#fff" stroke="#f97316" stroke-width="1.5"/>
+  <text x="209" y="244" font-size="11" text-anchor="middle" fill="#9a3412" font-weight="600">rag-mcp</text>
+  <text x="209" y="260" font-size="10" text-anchor="middle" fill="#94a3b8">MCP server</text>
+
+  <line x1="264" y1="248" x2="292" y2="248" stroke="#64748b" stroke-width="1.5" marker-end="url(#a)"/>
+
+  <!-- Embed query -->
+  <rect x="294" y="224" width="110" height="48" rx="6" fill="#fff" stroke="#22c55e" stroke-width="1.5"/>
+  <text x="349" y="244" font-size="11" text-anchor="middle" fill="#166534" font-weight="600">Embed query</text>
+  <text x="349" y="260" font-size="10" text-anchor="middle" fill="#94a3b8">query vector</text>
+
+  <line x1="404" y1="248" x2="432" y2="248" stroke="#64748b" stroke-width="1.5" marker-end="url(#a)"/>
+
+  <!-- HNSW search -->
+  <rect x="434" y="224" width="118" height="48" rx="6" fill="#fff" stroke="#22c55e" stroke-width="1.5"/>
+  <text x="493" y="244" font-size="11" text-anchor="middle" fill="#166534" font-weight="600">HNSW search</text>
+  <text x="493" y="260" font-size="10" text-anchor="middle" fill="#94a3b8">cosine distance</text>
+
+  <!-- DB → HNSW dashed -->
+  <line x1="712" y1="152" x2="712" y2="178" stroke="#94a3b8" stroke-width="1" stroke-dasharray="4,3"/>
+  <line x1="712" y1="178" x2="550" y2="224" stroke="#94a3b8" stroke-width="1" stroke-dasharray="4,3" marker-end="url(#ad)"/>
+
+  <line x1="552" y1="248" x2="580" y2="248" stroke="#64748b" stroke-width="1.5" marker-end="url(#a)"/>
+
+  <!-- Reranker -->
+  <rect x="582" y="224" width="118" height="48" rx="6" fill="#fff" stroke="#22c55e" stroke-width="1.5"/>
+  <text x="641" y="244" font-size="11" text-anchor="middle" fill="#166534" font-weight="600">Reranker</text>
+  <text x="641" y="260" font-size="10" text-anchor="middle" fill="#94a3b8">cross-encoder</text>
+
+  <!-- Reranker → Claude return arc -->
+  <path d="M700 262 Q760 296 720 296 Q440 296 140 296 Q114 296 72 276" stroke="#64748b" stroke-width="1.5" fill="none" marker-end="url(#a)"/>
+  <text x="430" y="309" font-size="10" text-anchor="middle" fill="#94a3b8">top results + symbol breadcrumbs</text>
+</svg>

package/docs/web-dashboard.md

@@ -49,7 +49,7 @@ can only reach the dashboard while the process is alive.
   `rag-config edit` (CLI) instead.
 - A failed **DB** connection is non-fatal: the config editor still works; the
   Projects and Graph tabs show an in-page error.
-- A failed **Ollama** connection only affects model suggestions (see
+- A failed **embedding backend** connection only affects model suggestions (see
   *Field options*); everything else works.
 
 ---
@@ -105,16 +105,15 @@ Some settings have a constrained set of valid values, so the form offers a list
 rather than free text. `GET /api/field-options` returns
 `{ [dottedKey]: { values: string[], strict: boolean } }`:
 
-- **`ollama.model` — strict.** The embedding model is tied to the vector column
-  dimension (`bge-m3` → 1024). Values come from `SUPPORTED_EMBED_MODELS`
-  (`src/constants.ts`); the dropdown does **not** allow free text, so a
-  dimension-incompatible model cannot be set. To add a model, confirm its
-  dimension matches and append it to `SUPPORTED_EMBED_MODELS`.
-- **`search.translateModel`, `reranker.model` — non-strict.** Suggestions are the
-  user's installed Ollama models (`ollama.list()`), but free text is allowed
-  because these are not dimension-bound and the user may pull a new model later.
-  If Ollama is unreachable the list is empty and the field degrades to plain
-  text.
+- **`embedding.model`, `reranker.model` — non-strict.** `embedding.model`
+  suggests `MODEL_REGISTRY` (`src/constants.ts`) entries plus any model the
+  backend currently serves; `reranker.model` suggests only served models. Both
+  allow free text — a model outside the registry still works (its dimension is
+  probed at index time). Served models come from `IEmbedder.listModels()`,
+  implemented only by `InfinityEmbedder` (2-second-limited `/models` listing);
+  if the backend is unreachable, or the provider doesn't implement `listModels`
+  (`openai-compatible`, `voyage`), the served-model list is empty and the field
+  degrades to registry suggestions only (or plain text for `reranker.model`).
 
 A strict `<select>` preserves an unknown current value as an extra option, so
 saving never silently drops a value that isn't in the list.
@@ -132,19 +131,39 @@ All routes are served from `127.0.0.1:<port>`.
 | `GET`  | `/api/defaults` | `CONFIG_DEFAULTS` — dotted key → default value |
 | `GET`  | `/api/field-options` | dotted key → `{ values, strict }` |
 | `GET`  | `/api/config` | Current config, `db.password` masked |
-| `PUT`  | `/api/config` | Save. Body = config JSON. `200 {ok:true}` or `400 {error}` |
+| `PUT`  | `/api/config` | Save. Body = config JSON. `200 {ok:true, backend:{action,message}}` or `400 {error}`. For `embedding.provider: infinity`, `backend.action` reports whether the Infinity container needed a background recreate (`started`/`already-running`/`not-needed`/`manual-required`/`failed`) — see *Infinity backend recreate* below |
+| `GET`  | `/api/backend-status` | `{ state, message, startedAt, finishedAt }` — progress of the most recent background Infinity recreate this server process started |
 | `GET`  | `/api/projects` | `SearcherService.listProjects()` |
 | `GET`  | `/api/graph` | `{ nodes, edges }` JSON (`buildGraphData()`) |
 | `GET`  | `/graph` | Full interactive graph page (`buildHtml()`), embedded by the Graph tab's iframe |
 
 ---
 
+## Infinity backend recreate
+
+When a save's `embedding.provider` is `infinity`, the server computes the
+required model set from the **saved** config (freshly re-read from disk, not
+the embedder the process started with) and compares it against what the
+backend currently serves. If a managed container needs to be started or
+recreated to serve the new set, that runs as an in-process background task —
+the `PUT` responds immediately (`action: "started"`) without waiting for it.
+Poll `GET /api/backend-status` for progress (`state: "recreating"` →
+`"ready"`/`"failed"`). Only one `docker compose` run is active at a time; a
+save that arrives while a recreate is already in flight for a managed backend
+responds `already-running`, and at most one follow-up run applies the latest
+such save once the in-flight one finishes. An external (unmanaged) backend
+missing a required model never triggers Docker — the response's `message`
+carries manual `--model-id`/launch instructions instead.
+
+---
+
 ## Architecture
 
 | File | Responsibility |
 |---|---|
 | `src/commands/WebCommand.ts` | CLI entry. `db.initDb()`, starts the server, keeps `execute()` pending until `SIGINT`/`SIGTERM` (mirrors `McpCommand`), then `server.stop()`. Parses `--port` (default `4319`). |
-| `src/services/WebServerService.ts` | `WebServer(searcher, config, logger)` — `node:http` server, routing, and the API handlers. Exports the pure `mergeConfigForSave()` and the `FieldOption` type. |
+| `src/services/WebServerService.ts` | `WebServer(searcher, config, logger, embedder, deps?)` — `node:http` server, routing, the API handlers, and the background Infinity recreate + `/api/backend-status` state. Exports the pure `mergeConfigForSave()` and the `FieldOption` type. |
+| `src/services/infinity-provisioning.ts` | The readiness/recreate primitive the background task delegates to (shared with `InfinityEmbedder` and `rag-setup`) — classification, compose-file rewrite, `docker compose` invocation, readiness polling. |
 | `src/web/dashboard.ts` | `renderDashboardHtml()` — the entire page (HTML + CSS + JS) as one self-contained string, mirroring `graph.ts:buildHtml`. |
 
 Reused primitives (no duplication):
@@ -153,7 +172,8 @@ Reused primitives (no duplication):
   `CONFIG_DEFAULTS` — `src/config-store.ts`
 - `buildGraphData` — `src/graph.ts`
 - `listProjects` — `src/services/SearcherService.ts`
-- `SUPPORTED_EMBED_MODELS` — `src/constants.ts`
+- `MODEL_REGISTRY` — `src/constants.ts`
+- `evaluateBackend` / `ensureModelSetServed` / `requiredModelSet` — `src/services/infinity-provisioning.ts`
 
 Wiring: `app.ts` detects the `web` mode and dispatches
 `new WebCommand(container, effectiveArgs)`. The command uses the shared
@@ -182,6 +202,5 @@ changes were needed.
   `.default(...)` calls, so the runtime default and the dashboard-displayed
   default cannot diverge — no manual sync needed. (A unit test still asserts
   every key is a valid config key.)
-- `SUPPORTED_EMBED_MODELS` (`constants.ts`) must contain only models whose
-  embedding dimension matches the `vector(...)` column — this one is still a
-  manual invariant, guarded only at the membership level by a unit test.
+- `MODEL_REGISTRY` (`constants.ts`) is advisory, not exhaustive — a model
+  outside it still works via the runtime dimension probe at index time.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "codesearch-mcp",
-  "version": "0.1.0-alpha",
+  "version": "0.2.1-alpha",
   "description": "Local semantic codebase search via MCP — indexes projects into pgvector, exposes search to Claude",
   "license": "MIT",
   "keywords": [
@@ -18,9 +18,6 @@
   "engines": {
     "node": ">=18"
   },
-  "publishConfig": {
-    "tag": "alpha"
-  },
   "main": "dist/app.js",
   "bin": {
     "rag-index": "dist/app.js",
@@ -45,7 +42,6 @@
     "start:web": "ts-node app.ts web",
     "index:workspace": "ts-node app.ts index",
     "index:reindex": "ts-node app.ts index --reindex",
-    "index:backfill": "ts-node app.ts backfill",
     "typecheck": "tsc --noEmit",
     "test": "vitest run",
     "test:watch": "vitest",
@@ -58,7 +54,6 @@
     "cli-progress": "^3.12.0",
     "drizzle-orm": "^0.45.2",
     "ignore": "^7.0.5",
-    "ollama": "^0.5.0",
     "pg": "^8.13.0",
     "pino": "^10.3.1",
     "pino-roll": "^4.0.0",

package/dist/src/commands/BackfillCommand.d.ts

@@ -1,7 +0,0 @@
-import type { Container } from '../container';
-import type { ICommand } from './ICommand';
-export declare class BackfillCommand implements ICommand {
-    private container;
-    constructor(container: Container);
-    execute(): Promise<void>;
-}

package/dist/src/commands/BackfillCommand.js

@@ -1,37 +0,0 @@
-"use strict";
-Object.defineProperty(exports, "__esModule", { value: true });
-exports.BackfillCommand = void 0;
-const interrupt_1 = require("../interrupt");
-class BackfillCommand {
-    container;
-    constructor(container) {
-        this.container = container;
-    }
-    async execute() {
-        const { db, indexer } = this.container;
-        await db.initDbWithSchema();
-        const doCalls = process.argv.includes('--calls');
-        const controller = new AbortController();
-        const teardown = (0, interrupt_1.setupInterrupt)(controller);
-        try {
-            if (doCalls) {
-                console.log('Backfilling call graph...\n');
-                const stats = await indexer.backfillCalls(controller.signal);
-                if (controller.signal.aborted)
-                    console.log('\nBackfill interrupted.');
-                console.log(`\nDone. Total: ${stats.total}, Processed: ${stats.processed}, Skipped (not on disk): ${stats.skipped}, Failed: ${stats.failed}`);
-            }
-            else {
-                console.log('Backfilling symbol index...\n');
-                const stats = await indexer.backfillSymbols(controller.signal);
-                if (controller.signal.aborted)
-                    console.log('\nBackfill interrupted.');
-                console.log(`\nDone. Total: ${stats.total}, Processed: ${stats.processed}, Skipped (not on disk): ${stats.skipped}, Failed: ${stats.failed}`);
-            }
-        }
-        finally {
-            teardown();
-        }
-    }
-}
-exports.BackfillCommand = BackfillCommand;

package/dist/src/services/OllamaEmbedder.d.ts

@@ -1,25 +0,0 @@
-import type { Config } from '../config';
-import type { IEmbedder, InputType } from '../interfaces';
-import type { ILogger } from '../interfaces';
-import type { EmbedderDeps } from './OpenAICompatibleEmbedder';
-export declare class OllamaEmbedder implements IEmbedder {
-    private config;
-    private logger;
-    readonly dimensions: number;
-    private client;
-    rerank?: (query: string, documents: string[], topN: number) => Promise<number[]>;
-    constructor(config: Config, logger?: ILogger, deps?: EmbedderDeps);
-    embed(text: string, _inputType?: InputType): Promise<number[]>;
-    embedBatch(texts: string[], _inputType?: InputType): Promise<number[][]>;
-    listModels(): Promise<string[]>;
-    /**
-     * Pull the configured model if it is not already present locally, so a plain
-     * `rag-index` works after switching to (or back to) a local model that was
-     * never pulled — `rag-setup` is no longer the only place that fetches models.
-     * A tagless model name (`bge-m3`) matches its `:latest`/tagged install.
-     * Streams pull progress to the console: a live single-line bar on a TTY, or
-     * de-duplicated status lines otherwise.
-     */
-    ensureModelAvailable(): Promise<void>;
-    private static formatPullProgress;
-}

package/dist/src/services/OllamaEmbedder.js

@@ -1,111 +0,0 @@
-"use strict";
-Object.defineProperty(exports, "__esModule", { value: true });
-exports.OllamaEmbedder = void 0;
-const ollama_1 = require("ollama");
-const LoggerService_1 = require("./LoggerService");
-const reranker_1 = require("./reranker");
-const constants_1 = require("../constants");
-class OllamaEmbedder {
-    config;
-    logger;
-    dimensions;
-    client;
-    rerank;
-    constructor(config, logger = LoggerService_1.LoggerService.silent(), deps = {}) {
-        this.config = config;
-        this.logger = logger;
-        this.client = new ollama_1.Ollama({ host: config.embedding.baseURL });
-        this.dimensions = config.embedding.dimensions ?? (0, constants_1.registryDimension)(config.embedding.model) ?? 0;
-        this.rerank = (0, reranker_1.makeReranker)(config, logger, deps);
-    }
-    // Ollama has no query/document asymmetry — inputType is accepted and ignored.
-    async embed(text, _inputType) {
-        const { embedding } = await this.client.embeddings({
-            model: this.config.embedding.model,
-            prompt: text,
-        });
-        return embedding;
-    }
-    async embedBatch(texts, _inputType) {
-        if (texts.length === 0)
-            return [];
-        try {
-            // /api/embed accepts an array input and returns one vector per input.
-            const res = await this.client.embed({ model: this.config.embedding.model, input: texts });
-            if (res.embeddings?.length === texts.length)
-                return res.embeddings;
-            this.logger.warn('OllamaEmbedder: batch embed count mismatch, falling back to sequential', {
-                expected: texts.length, got: res.embeddings?.length ?? 0,
-            });
-        }
-        catch (err) {
-            this.logger.warn('OllamaEmbedder: batch embed failed, falling back to sequential', {
-                reason: err instanceof Error ? err.message : String(err),
-            });
-        }
-        // Sequential fallback — the contract (one vector per input, in order) holds.
-        const out = [];
-        for (const t of texts)
-            out.push(await this.embed(t));
-        return out;
-    }
-    async listModels() {
-        const res = await this.client.list();
-        return res.models.map(m => m.name).filter(Boolean).sort();
-    }
-    /**
-     * Pull the configured model if it is not already present locally, so a plain
-     * `rag-index` works after switching to (or back to) a local model that was
-     * never pulled — `rag-setup` is no longer the only place that fetches models.
-     * A tagless model name (`bge-m3`) matches its `:latest`/tagged install.
-     * Streams pull progress to the console: a live single-line bar on a TTY, or
-     * de-duplicated status lines otherwise.
-     */
-    async ensureModelAvailable() {
-        const model = this.config.embedding.model;
-        try {
-            const res = await this.client.list();
-            const present = res.models.some(m => m.name === model || m.name.startsWith(`${model}:`));
-            if (present)
-                return;
-        }
-        catch (err) {
-            // Can't list (Ollama unreachable?) — try the pull; its error is clearer.
-            this.logger.warn('OllamaEmbedder: could not list models before pull', {
-                reason: err instanceof Error ? err.message : String(err),
-            });
-        }
-        console.log(`Pulling Ollama model "${model}" (not found locally)...`);
-        this.logger.info('OllamaEmbedder: pulling model', { model });
-        const stream = await this.client.pull({ model, stream: true });
-        const isTty = Boolean(process.stdout.isTTY);
-        let lastLen = 0;
-        let lastLog = '';
-        for await (const part of stream) {
-            const out = '  ' + OllamaEmbedder.formatPullProgress(part);
-            if (isTty) {
-                process.stdout.write('\r' + out.padEnd(lastLen));
-                lastLen = out.length;
-            }
-            else if (out !== lastLog) {
-                console.log(out);
-                lastLog = out;
-            }
-        }
-        if (isTty && lastLen > 0)
-            process.stdout.write('\r' + ' '.repeat(lastLen) + '\r');
-        console.log(`  ✓ ${model} ready.`);
-    }
-    static formatPullProgress(p) {
-        const status = p.status ?? 'pulling';
-        if (p.total && p.total > 0) {
-            const pct = Math.min(100, Math.floor(((p.completed ?? 0) / p.total) * 100));
-            const width = 24;
-            const filled = Math.round((pct / 100) * width);
-            const bar = '[' + '█'.repeat(filled) + '░'.repeat(width - filled) + ']';
-            return `${status} ${bar} ${pct}% (${((p.completed ?? 0) / 1e6).toFixed(1)}/${(p.total / 1e6).toFixed(1)} MB)`;
-        }
-        return status;
-    }
-}
-exports.OllamaEmbedder = OllamaEmbedder;