System Overview

The primary product — a Model Context Protocol server that AI agents connect to for tool discovery. Exposes five MCP tools and orchestrates the search package.

Next.js admin dashboard for monitoring graph health, managing tool data, and viewing analytics.

Next.js public-facing site with documentation, search playground, and marketing pages.

GitHub indexer that scans repositories on a schedule, extracts tool metadata, and queues update jobs for workers.

Shared TypeScript types, Zod schemas, and domain constants used across all apps and packages.

Memgraph client and Cypher query builders. Provides repository interfaces for graph reads and writes over the Bolt protocol.

Qdrant client, Nomic Embed Code 768-dimensional embeddings, and similarity search helpers.

The 4-stage search pipeline — BM25 + vector retrieval, payload filtering, graph reranking, and selection logic.

PostgreSQL access via Prisma. Handles session storage, analytics events, and user preferences.

Redis Streams wrapper (ioredis). Provides reliable job queuing for the indexer worker pipeline.

Validated configuration module — Zod schema over process.env. Single source of truth for all environment variables.

Memgraph

Native graph processing engine. Tool relationships (RELATED_TO, DEPENDS_ON, ALTERNATIVE_TO) are first-class graph edges, making traversal and reranking queries orders of magnitude faster than join-heavy SQL.

Qdrant

High-performance vector similarity search with rich payload filtering. Supports hybrid BM25 + dense vector queries in a single request, which maps directly to Stage 1 of the search pipeline.

PostgreSQL

ACID-compliant relational storage for session data, analytics events, and user preferences — workloads that benefit from strong consistency and mature tooling (Prisma).

Redis Streams

Reliable, append-only job queue with consumer groups. Ensures indexer workers process every update exactly once, even across restarts, without the operational weight of a dedicated message broker.