mozempk/trueref
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(TRUEREF-0023): add sqlite-vec search pipeline

Browse Source
This commit is contained in:
Giancarmine Salucci
2026-04-01 14:09:19 +02:00
parent 0752636847
commit 9525c58e9a
45 changed files with 4009 additions and 614 deletions
Download Patch File Download Diff File Expand all files Collapse all files

189
docs/ARCHITECTURE.md
View File

@@ -1,171 +1,168 @@
# Architecture
Last Updated: 2026-03-30T00:00:00.000Z
Last Updated: 2026-04-01T12:05:23.000Z
## Overview
TrueRef is a TypeScript-first, self-hosted documentation retrieval platform built on SvelteKit. The repository contains a Node-targeted web application, a REST API, a Model Context Protocol server, and a multi-threaded server-side indexing pipeline backed by SQLite via better-sqlite3 and Drizzle ORM.
TrueRef is a TypeScript-first, self-hosted documentation retrieval platform built on SvelteKit. The repository contains a Node-targeted web application, a REST API, a Model Context Protocol server, and a worker-threaded indexing pipeline backed by SQLite via better-sqlite3, Drizzle ORM, FTS5, and sqlite-vec.
- Primary language: TypeScript (141 files) with a small amount of JavaScript configuration (2 files)
- Application type: Full-stack SvelteKit application with worker-threaded indexing and retrieval services
- Primary language: TypeScript (147 `.ts` files) with a small amount of JavaScript configuration and build code (2 `.js` files), excluding generated output and dependencies
- Application type: Full-stack SvelteKit application with server-side indexing, retrieval, and MCP integration
- Runtime framework: SvelteKit with adapter-node
- Storage: SQLite (WAL mode) with Drizzle-managed schema plus hand-written FTS5 setup
- Concurrency: Node.js worker_threads for parse and embedding work
- Testing: Vitest with separate client and server projects
- Storage: SQLite in WAL mode with Drizzle-managed relational schema, FTS5 full-text indexes, and sqlite-vec virtual tables for vector lookup
- Concurrency: Node.js `worker_threads` for parse, embed, and auxiliary write-worker infrastructure
- Testing: Vitest for unit and integration coverage
## Project Structure
- src/routes: SvelteKit pages and HTTP endpoints, including the public UI and /api/v1 surface
- src/lib/server: Backend implementation grouped by concern: api, config, crawler, db, embeddings, mappers, models, parser, pipeline, search, services, utils
- src/mcp: Standalone MCP server entry point and tool handlers
- static: Static assets such as robots.txt
- docs/features: Feature-level implementation notes and product documentation
- build: Generated SvelteKit output
- `src/routes`: SvelteKit pages and HTTP endpoints, including the public UI and `/api/v1` surface
- `src/lib/server`: Backend implementation grouped by concern: `api`, `config`, `crawler`, `db`, `embeddings`, `mappers`, `models`, `parser`, `pipeline`, `search`, `services`, `utils`
- `src/mcp`: Standalone MCP server entry point, client, tests, and tool handlers
- `scripts`: Build helpers, including worker bundling
- `static`: Static assets such as `robots.txt`
- `docs/features`: Feature-level implementation notes and product documentation
- `build`: Generated SvelteKit output and bundled worker entrypoints
## Key Directories
### src/routes
### `src/routes`
Contains the UI entry points and API routes. The API tree under src/routes/api/v1 is the public HTTP contract for repository management, indexing jobs, search/context retrieval, settings, filesystem browsing, JSON schema discovery, real-time SSE progress streaming, and job control (pause/resume/cancel).
Contains the UI entry points and API routes. The API tree under `src/routes/api/v1` is the public HTTP contract for repository management, version discovery, indexing jobs, search/context retrieval, embedding settings, indexing settings, filesystem browsing, worker-status inspection, and SSE progress streaming.
### src/lib/server/db
### `src/lib/server/db`
Owns SQLite schema definitions, migration bootstrapping, and FTS initialization. Database startup runs through initializeDatabase(), which executes Drizzle migrations and then applies FTS5 SQL that cannot be expressed directly in the ORM.
Owns SQLite schema definitions, relational migrations, connection bootstrapping, and sqlite-vec loading. Database startup goes through `initializeDatabase()` and `getClient()`, both of which configure WAL-mode pragmas and ensure sqlite-vec is loaded on each connection before vector-backed queries run.
### src/lib/server/pipeline
### `src/lib/server/search`
Coordinates crawl, parse, chunk, store, and optional embedding generation work using a worker thread pool. The pipeline module consists of:
Implements keyword, vector, and hybrid retrieval. Keyword search uses SQLite FTS5 and BM25-style ranking. Vector search uses `SqliteVecStore` to maintain per-profile sqlite-vec `vec0` tables plus rowid mapping tables, and hybrid search blends FTS and vector candidates through reciprocal rank fusion.
- **WorkerPool** (`worker-pool.ts`): Manages a configurable number of Node.js `worker_threads` for parse jobs and an optional dedicated embed worker. Dispatches jobs round-robin to idle workers, enforces per-repository serialisation (one active job per repo), auto-respawns crashed workers, and supports runtime concurrency adjustment via `setMaxConcurrency()`. Falls back to main-thread execution when worker scripts are not found.
- **Parse worker** (`worker-entry.ts`): Runs in a worker thread. Opens its own `better-sqlite3` connection (WAL mode, `busy_timeout = 5000`), constructs a local `IndexingPipeline` instance, and processes jobs by posting `progress`, `done`, or `failed` messages back to the parent.
- **Embed worker** (`embed-worker-entry.ts`): Dedicated worker for embedding generation. Loads the embedding profile from the database, creates an `EmbeddingService`, and processes embed requests after the parse worker finishes a job.
- **ProgressBroadcaster** (`progress-broadcaster.ts`): Server-side pub/sub for real-time SSE streaming. Supports per-job, per-repository, and global subscriptions. Caches the last event per job for reconnect support.
- **Worker types** (`worker-types.ts`): Shared TypeScript discriminated union types for `ParseWorkerRequest`/`ParseWorkerResponse` and `EmbedWorkerRequest`/`EmbedWorkerResponse` message protocols.
- **Startup** (`startup.ts`): Recovers stale jobs, constructs singleton `JobQueue`, `IndexingPipeline`, `WorkerPool`, and `ProgressBroadcaster` instances, reads concurrency settings from the database, and drains queued work after restart.
- **JobQueue** (`job-queue.ts`): SQLite-backed queue that delegates to the `WorkerPool` when available, with pause/resume/cancel support.
### `src/lib/server/pipeline`
### src/lib/server/search
Coordinates crawl, diff, parse, store, embed, and job-state broadcasting. The pipeline module consists of:
Implements keyword, vector, and hybrid retrieval. The keyword path uses SQLite FTS5 and BM25; the hybrid path blends FTS and vector search with reciprocal rank fusion.
- `IndexingPipeline`: orchestrates crawl, diff, parse, transactional replacement, optional embedding generation, and repository statistics updates
- `WorkerPool`: manages parse workers, an optional embed worker, an optional write worker, per-repository-and-version serialization, worker respawn, and runtime concurrency changes
- `worker-entry.ts`: parse worker that opens its own `better-sqlite3` connection, runs the indexing pipeline, and reports progress back to the parent
- `embed-worker-entry.ts`: embedding worker that loads the active profile, creates an `EmbeddingService`, and generates vectors after parse completion
- `write-worker-entry.ts`: batch-write worker with a `write`/`write_ack`/`write_error` message protocol for document and snippet persistence
- `progress-broadcaster.ts`: server-side pub/sub for per-job, per-repository, global, and worker-status SSE streams
- `startup.ts`: recovers stale jobs, constructs singleton queue/pipeline/pool/broadcaster instances, loads concurrency settings, and drains queued work after restart
- `worker-types.ts`: shared TypeScript discriminated unions for parse, embed, and write worker protocols
### src/lib/server/crawler and src/lib/server/parser
### `src/lib/server/crawler` and `src/lib/server/parser`
Convert GitHub repositories and local folders into normalized snippet records. Crawlers fetch repository contents, parsers split Markdown, code, config, HTML-like, and plain-text files into chunks, and downstream services persist searchable content.
Convert GitHub repositories and local folders into normalized snippet records. Crawlers fetch repository contents and configuration, parsers split Markdown, code, config, HTML-like, and plain-text files into searchable snippet records, and downstream services persist searchable content and embeddings.
### src/mcp
### `src/mcp`
Provides a thin compatibility layer over the HTTP API. The MCP server exposes resolve-library-id and query-docs over stdio or HTTP and forwards work to local tool handlers.
Provides a thin compatibility layer over the HTTP API. The MCP server exposes `resolve-library-id` and `query-docs` over stdio or HTTP and forwards work to local handlers that reuse the application retrieval stack.
## Design Patterns
- The WorkerPool implements an **observer/callback pattern**: the pool owner provides `onProgress`, `onJobDone`, `onJobFailed`, `onEmbedDone`, and `onEmbedFailed` callbacks at construction time, and the pool invokes them when workers post messages.
- ProgressBroadcaster implements a **pub/sub pattern** with three subscription tiers (per-job, per-repository, global) and last-event caching for SSE reconnect.
- The implementation consistently uses **service classes** such as RepositoryService, SearchService, and HybridSearchService for business logic.
- Mapping and entity layers separate raw database rows from domain objects through **mapper/entity pairs** such as RepositoryMapper and RepositoryEntity.
- Pipeline startup uses **module-level singletons** for JobQueue, IndexingPipeline, WorkerPool, and ProgressBroadcaster lifecycle management, with accessor functions (getQueue, getPool, getBroadcaster) for route handlers.
- Worker message protocols use **TypeScript discriminated unions** (`type` field) for type-safe worker ↔ parent communication.
- **Service layer**: business logic lives in classes such as `RepositoryService`, `VersionService`, `SearchService`, `HybridSearchService`, and `EmbeddingService`
- **Factory pattern**: embedding providers are created from persisted profile records through registry/factory helpers
- **Mapper/entity separation**: mappers translate between raw database rows and domain entities such as `RepositoryEntity`, `RepositoryVersionEntity`, and `EmbeddingProfileEntity`
- **Module-level singletons**: pipeline startup owns lifecycle for `JobQueue`, `IndexingPipeline`, `WorkerPool`, and `ProgressBroadcaster`, with accessor functions for route handlers
- **Pub/sub**: `ProgressBroadcaster` maintains job, repository, global, and worker-status subscriptions for SSE delivery
- **Discriminated unions**: worker message protocols use a `type` field for type-safe parent/worker communication
## Key Components
### SvelteKit server bootstrap
src/hooks.server.ts initializes the database, loads persisted embedding configuration, creates the optional EmbeddingService, reads indexing concurrency settings from the database, starts the indexing pipeline with WorkerPool and ProgressBroadcaster via `initializePipeline(db, embeddingService, { concurrency, dbPath })`, and applies CORS headers to all /api routes.
`src/hooks.server.ts` initializes the relational database, opens the shared raw SQLite client, loads the default embedding profile, creates the optional `EmbeddingService`, reads indexing concurrency from the `settings` table, and initializes the queue/pipeline/worker infrastructure.
### Database layer
src/lib/server/db/schema.ts defines repositories, repository_versions, documents, snippets, embedding_profiles, snippet_embeddings, indexing_jobs, repository_configs, and settings. This schema models the indexed library catalog, retrieval corpus, embedding state, and job tracking.
`src/lib/server/db/schema.ts` defines repositories, repository versions, documents, snippets, embedding profiles, relational embedding metadata, indexing jobs, repository configs, and generic settings. Relational embedding rows keep canonical model metadata and raw float buffers, while sqlite-vec virtual tables are managed separately per profile through `SqliteVecStore`.
### sqlite-vec integration
`src/lib/server/db/sqlite-vec.ts` centralizes sqlite-vec loading and deterministic per-profile table naming. `SqliteVecStore` creates `vec0` tables plus rowid mapping tables, backfills missing rows from `snippet_embeddings`, removes stale vector references, and executes nearest-neighbor queries constrained by repository, optional version, and profile.
### Retrieval API
src/routes/api/v1/context/+server.ts validates input, resolves repository and optional version IDs, chooses keyword, semantic, or hybrid retrieval, applies token budgeting that skips oversized snippets instead of stopping early, prepends repository rules, and formats JSON or text responses with repository and version metadata.
`src/routes/api/v1/context/+server.ts` validates input, resolves repository and optional version scope, chooses keyword, semantic, or hybrid retrieval, applies token budgeting, and formats JSON or text responses. `/api/v1/libs/search` handles repository-level lookup, while MCP tool handlers expose the same retrieval behavior over stdio or HTTP transports.
### Search engine
src/lib/server/search/search.service.ts preprocesses raw user input into FTS5-safe MATCH expressions before keyword search and repository lookup. src/lib/server/search/hybrid.search.service.ts supports explicit keyword, semantic, and hybrid modes, falls back to vector retrieval when FTS yields no candidates and an embedding provider is configured, and uses reciprocal rank fusion for blended ranking.
`SearchService` preprocesses raw user input into FTS5-safe expressions before keyword search. `HybridSearchService` supports explicit keyword, semantic, and hybrid modes, falls back to vector retrieval when keyword search yields no candidates and an embedding provider is configured, and uses reciprocal rank fusion to merge ranked lists. `VectorSearch` delegates KNN execution to `SqliteVecStore` instead of doing brute-force in-memory cosine scoring.
### Repository management
### Repository and version management
src/lib/server/services/repository.service.ts provides CRUD and statistics for indexed repositories, including canonical ID generation for GitHub and local sources.
`RepositoryService` and `VersionService` provide CRUD, indexing-status, cleanup, and statistics logic for indexed repositories and tagged versions, including sqlite-vec cleanup when repository-scoped or version-scoped content is removed.
### MCP surface
### Worker-threaded indexing
src/mcp/index.ts creates the MCP server, registers the two supported tools, and exposes them over stdio or streamable HTTP.
The active indexing path is parse-worker-first: queued jobs are dispatched to parse workers, progress is written to SQLite and broadcast over SSE, and successful parse completion can enqueue embedding work on the dedicated embed worker. The worker pool also exposes status snapshots through `/api/v1/workers`. Write-worker infrastructure exists in the current architecture and is bundled at build time, but parse/embed flow remains the primary live path described by `IndexingPipeline` and `WorkerPool`.
### Worker thread pool
### SSE streaming and job control
src/lib/server/pipeline/worker-pool.ts manages a pool of Node.js worker threads. Parse workers run the full crawl → parse → store pipeline inside isolated threads with their own better-sqlite3 connections (WAL mode enables concurrent readers). An optional embed worker handles embedding generation in a separate thread. The pool enforces per-repository serialisation, auto-respawns crashed workers, and supports runtime concurrency changes persisted through the settings table.
### SSE streaming
src/lib/server/pipeline/progress-broadcaster.ts provides real-time Server-Sent Event streaming of indexing progress. Route handlers in src/routes/api/v1/jobs/stream and src/routes/api/v1/jobs/[id]/stream expose SSE endpoints. The broadcaster supports per-job, per-repository, and global subscriptions, with last-event caching for reconnect via the `Last-Event-ID` header.
### Job control
src/routes/api/v1/jobs/[id]/pause, resume, and cancel endpoints allow runtime control of indexing jobs. The JobQueue supports pause/resume/cancel state transitions persisted to SQLite.
`progress-broadcaster.ts` provides real-time Server-Sent Event streaming of indexing progress. Route handlers under `/api/v1/jobs/stream` and `/api/v1/jobs/[id]/stream` expose SSE endpoints, and `/api/v1/workers` exposes worker-pool status. Job control endpoints support pause, resume, and cancel transitions backed by SQLite job state.
### Indexing settings
src/routes/api/v1/settings/indexing exposes GET and PUT for indexing concurrency. PUT validates and clamps the value to `max(cpus - 1, 1)`, persists it to the settings table, and live-updates the WorkerPool via `setMaxConcurrency()`.
`/api/v1/settings/indexing` exposes GET and PUT for indexing concurrency. The value is persisted in the `settings` table and applied live to the `WorkerPool` through `setMaxConcurrency()`.
## Dependencies
### Production
- @modelcontextprotocol/sdk: MCP server transport and protocol types
- @xenova/transformers: local embedding support
- better-sqlite3: synchronous SQLite driver
- zod: runtime input validation for MCP tools and server helpers
- `@modelcontextprotocol/sdk`: MCP server transport and protocol types
- `@xenova/transformers`: local embedding support
- `better-sqlite3`: synchronous SQLite driver used by the main app and workers
- `sqlite-vec`: SQLite vector extension used for `vec0` storage and nearest-neighbor queries
- `zod`: runtime validation for MCP tools and server helpers
### Development
- @sveltejs/kit and @sveltejs/adapter-node: application framework and Node deployment target
- drizzle-kit and drizzle-orm: schema management and typed database access
- esbuild: worker thread entry point bundling (build/workers/)
- vite and @tailwindcss/vite: bundling and Tailwind integration
- vitest and @vitest/browser-playwright: server and browser test execution
- eslint, typescript-eslint, eslint-plugin-svelte, prettier, prettier-plugin-svelte, prettier-plugin-tailwindcss: linting and formatting
- typescript and @types/node: type-checking and Node typings
- `@sveltejs/kit` and `@sveltejs/adapter-node`: application framework and Node deployment target
- `drizzle-kit` and `drizzle-orm`: schema management and typed database access
- `esbuild`: worker entrypoint bundling into `build/workers`
- `vite` and `@tailwindcss/vite`: application bundling and Tailwind integration
- `vitest` and `@vitest/browser-playwright`: server and browser test execution
- `eslint`, `typescript-eslint`, `eslint-plugin-svelte`, `prettier`, `prettier-plugin-svelte`, `prettier-plugin-tailwindcss`: linting and formatting
- `typescript` and `@types/node`: type-checking and Node typings
## Module Organization
The backend is organized by responsibility rather than by route. HTTP handlers in src/routes/api/v1 are intentionally thin and delegate to library modules in src/lib/server. Within src/lib/server, concerns are separated into:
The backend is organized by responsibility rather than by route. HTTP handlers under `src/routes/api/v1` are intentionally thin and delegate to modules in `src/lib/server`. Within `src/lib/server`, concerns are separated into:
- models and mappers for entity translation
- services for repository/version operations
- search for retrieval strategies
- crawler and parser for indexing input transformation
- pipeline for orchestration and job execution
- embeddings for provider abstraction and embedding generation
- api and utils for response formatting, validation, and shared helpers
- `models` and `mappers` for entity translation
- `services` for repository/version operations
- `search` for keyword, vector, and hybrid retrieval strategies
- `crawler` and `parser` for indexing input transformation
- `pipeline` for orchestration, workers, and job execution
- `embeddings` for provider abstraction and vector generation
- `db`, `api`, `config`, and `utils` for persistence, response formatting, validation, and shared helpers
The frontend and backend share the same SvelteKit repository, but most non-UI behavior is implemented on the server side.
The frontend and backend live in the same SvelteKit repository, but most non-UI behavior is implemented on the server side.
## Data Flow
### Indexing flow
1. Server startup runs initializeDatabase() and initializePipeline() from src/hooks.server.ts, which creates the WorkerPool, ProgressBroadcaster, and JobQueue singletons.
2. The pipeline recovers stale jobs (marks running → failed, indexing → error), reads concurrency settings, and resumes queued work.
3. When a job is enqueued, the JobQueue delegates to the WorkerPool, which dispatches work to an idle parse worker thread.
4. Each parse worker opens its own better-sqlite3 connection (WAL mode) and runs the full crawl → parse → store pipeline, posting progress messages back to the parent thread.
5. The parent thread updates job progress in the database and broadcasts SSE events through the ProgressBroadcaster.
6. On parse completion, if an embedding provider is configured, the WorkerPool enqueues an embed request to the dedicated embed worker, which generates vectors in its own thread.
7. Job control endpoints allow pausing, resuming, or cancelling jobs at runtime.
1. Server startup runs database initialization, opens the shared client, loads sqlite-vec, and initializes the pipeline singletons.
2. Startup recovery marks interrupted jobs as failed, resets repositories stuck in `indexing`, reads persisted concurrency settings, and drains queued jobs.
3. `JobQueue` dispatches eligible work to the `WorkerPool`, which serializes by `(repositoryId, versionId)` and posts jobs to idle parse workers.
4. Each parse worker opens its own SQLite connection, crawls the source, computes differential work, parses files into snippets, and persists replacement data through the indexing pipeline.
5. The parent thread updates job progress in SQLite and broadcasts SSE progress and worker-status events.
6. If an embedding provider is configured, the completed parse job triggers embed work that stores canonical embedding blobs and synchronizes sqlite-vec profile tables for nearest-neighbor lookup.
7. Repository/version statistics and job status are finalized in SQLite, and control endpoints can pause, resume, or cancel subsequent queued work.
### Retrieval flow
1. Clients call /api/v1/libs/search, /api/v1/context, or the MCP tools.
2. Route handlers validate input and load the SQLite client.
3. Keyword search uses FTS5 via SearchService; hybrid search optionally adds vector results via HybridSearchService.
4. Query preprocessing normalizes punctuation-heavy or code-like input before FTS search, while semantic mode bypasses FTS and auto or hybrid mode can fall back to vector retrieval when keyword search produces no candidates.
5. Token budgeting walks ranked snippets in order and skips individual over-budget snippets so later matches can still be returned.
6. Formatters emit repository and version metadata in JSON responses and origin-aware or explicit no-result text output for plain-text responses.
7. MCP handlers expose the same retrieval behavior over stdio or HTTP transports.
1. Clients call `/api/v1/libs/search`, `/api/v1/context`, or the MCP tools.
2. Route handlers validate input and use the shared SQLite client.
3. Keyword search uses FTS5 through `SearchService`; semantic search uses sqlite-vec KNN through `VectorSearch`; hybrid search merges both paths with reciprocal rank fusion.
4. Retrieval is scoped by repository and optional version, and semantic/hybrid paths can fall back when keyword search yields no usable candidates.
5. Token budgeting selects ranked snippets for the response formatter, which emits repository-aware JSON or text payloads.
## Build System
- Build command: npm run build (runs `vite build` then `node scripts/build-workers.mjs`)
- Worker bundling: scripts/build-workers.mjs uses esbuild to compile worker-entry.ts and embed-worker-entry.ts into build/workers/ as ESM bundles (.mjs), with $lib path aliases resolved and better-sqlite3/@xenova/transformers marked external
- Test command: npm run test
- Primary local run command from package.json: npm run dev
- MCP entry points: npm run mcp:start and npm run mcp:http
- Build command: `npm run build` (runs `vite build` then `node scripts/build-workers.mjs`)
- Worker bundling: `scripts/build-workers.mjs` uses esbuild to compile `worker-entry.ts`, `embed-worker-entry.ts`, and `write-worker-entry.ts` into `build/workers/` as ESM bundles
- Test command: `npm test`
- Primary local run command: `npm run dev`
- MCP entry points: `npm run mcp:start` and `npm run mcp:http`