xb

AI Application Starter (English)

This page consolidates guidance that used to live under xb/doc/ai_application/. It focuses on three workflows: Retrieval-Augmented Generation (RAG), Agent tooling, and analytics dashboards backed by xb.

---

1. Retrieval-Augmented Generation (RAG)

1.1 Data pipeline checklist

1. Chunking – split documents into 512-1024 token windows.
2. Embedding – store the raw vector, the text, and metadata (tenant_id, lang, tags).
3. Indexing – choose a Qdrant collection per tenant or per domain.
4. Sync metadata – keep SQL + vector stores in lockstep via CDC or async jobs.

1.2 Query pipeline

queryVec := embedder.Encode(prompt)

resultsJSON, _ := xb.Of(&DocVector{}).
    Custom(qdrantCustom).
    Eq("tenant_id", tenantID).
    VectorSearch("embedding", queryVec, 8).
    Build().
    JsonOfSelect()

Feed resultsJSON to the RAG orchestrator (LangChain, LlamaIndex, Semantic Kernel, etc.). Each toolkit consumes the same payload because xb sticks to raw Qdrant JSON.

---

2. Agent tooling

• Tool schema – expose builder presets as agent tools (recommend_feed, discover_code, scroll_notifications).
• Guard rails – use Meta(func) to inject agent/session IDs and enforce quotas in interceptors.
• Streaming – for long-running scrolls, paginate via ScrollID and send chunks to the agent runtime.

Example tool definition:

func RecommendFeedTool(input RecommendInput) (string, error) {
    json, err := xb.Of(&FeedVector{}).
        Custom(qdrantCustom).
        Eq("tenant_id", input.Tenant).
        VectorSearch("embedding", input.Vector, input.Limit).
        Build().
        JsonOfSelect()
    if err != nil {
        return "", err
    }
    return string(json), nil
}

---

3. Analytics dashboards

Even if the front-end is SQL-only, xb helps keep AI activity auditable:

• Use Meta(func) to log TraceID, UserID, Model per query.
• Register AfterBuild interceptors to push query summaries to observability systems.
• Store built.SqlOfSelect() alongside built.JsonOfSelect() to correlate SQL + vector traces.

---

4. Integration notes

Toolkit	Tip
LangChain	Wrap xb builders as Runnable objects; use JsonOfSelect output directly
LlamaIndex	Register a custom retriever that calls xb; keep metadata consistent
Semantic Kernel	Implement IQueryFunction to invoke xb and return embeddings

---

5. Where to go next

• QUICKSTART.md for the core builder API
• QDRANT_GUIDE.md for advanced vector payloads
• VECTOR_GUIDE.md for embedding hygiene and diversity settings
• FILTERING.md to understand how inputs are sanitized automatically

If you have a repeatable AI workflow (RAG, agent, analytics) worth sharing, PRs to this file are welcome.

This site is open source. Improve this page.