Agents Frame is a four-stage pipeline. Every entry point (REST, MCP, CLI) funnels into the same engine, so behavior is identical regardless of how you call in. 
intent ──► detectLang ──► embed ──► recall ──► rerank ──► selected
                (en/zh)   (Gemini)  (pgvector  (ELO-       framework
                          1536d)    cosine)    weighted)   + promptMd

​
1. Language detection

The router auto-detects whether your intent is English or Chinese (the two languages currently supported). You can override with the lang field. Every framework has a parallel bilingual prompt, so switching language switches the whole experience — not just the surface text.

​
2. Embedding

Intents are embedded with Google’s gemini-embedding-001 model at 1536 dimensions. Embeddings live close to the database (Vertex AI in the same region as Neon Postgres us-east-1) to keep round-trip latency in the low hundreds of milliseconds.

​
3. Recall

The embedded intent is matched against FrameworkEmbedding rows in Postgres using pgvector cosine similarity. We return the top recallK frameworks (default 16) — wide enough to leave room for the re-ranker, narrow enough to stay cheap.
pgvector is the only vector store in the loop. There’s no separate ANN service, no embedding cache layer, no third-party retrieval hop. Fewer moving parts, fewer places to debug.

​
4. Rerank

Each candidate carries an ELO score updated from real user feedback (UserFeedback table). The final score blends cosine similarity with the framework’s win-rate, so frameworks that actually help float to the top over time. The top-1 candidate becomes selected; the rest are returned as candidates so you can inspect alternatives or build your own selection UI.

​
Output shape

type RouteResult = {
  selected: {
    slug: string;
    frameworkVersionId: string;  // pin this for reproducibility
    promptMd: string;            // the actual prompt to use
  };
  candidates: Array<{ slug: string; score: number }>;
  detectedLang: 'en' | 'zh';
  latencyMs: number;
};

​
Stateless by design

The router holds no conversation state. Each /v1/think call is a pure function of (intent, lang, topK, recallK, category, excludeSlugs) plus the framework library version. Your agent keeps its own memory; Agents Frame just answers a single question — which framework? — and answers it the same way every time.

​
What’s next

The same pipeline will scale from 5 frameworks to 30 without any contract change. Future versions will add categories (decision, debugging, planning, communication), cross-lingual recall, and optional RAG-style enrichment from your private framework set.