Skip to content

AI in Dédalo

The AI subsystem — semantic search, grounded Q&A, object-image similarity, and the in-app assistant / agent / MCP surface. All of it is a greenfield TypeScript/Bun build on top of the stable typed core, strictly opt-in, and gated by the same permissions (ACL) as human access: an AI query returns exactly what the same user could read by hand — never more.

New to the vocabulary (section, component, tipo, locator, sqo, rqo, ddo)? Start with the core hub and the Glossary.

The two layers

Dédalo's AI is built in two layers — a retrieval/generation engine and the surfaces that consume it:

flowchart TB
    subgraph SURF["surfaces"]
        CHAT["In-app Assistant<br/>(chat UI)"]
        AGENT["Agent loop<br/>(tool-calling)"]
        MCP["MCP server<br/>(stdio / HTTP)"]
        CHAT ~~~ AGENT ~~~ MCP
    end

    SURF -->|"same ACL-gated handlers<br/>resolveCaller · getPermissions"| ENG

    subgraph ENG["engine — retrieval / generation"]
        direction TB
        SS["semantic search"]
        HR["hybrid retrieval<br/>dense + lexical, RRF fusion"]
        GA["grounded ask<br/>retrieval → LLM, cited"]
        OI["object-image similarity<br/>& characterization"]
        RAG["RAG<br/>dd_rag_api · src/ai/rag/"]
        RAG --> SS
        RAG --> HR
        RAG --> GA  
        RAG --> OI
    end
  • RAG & Semantic Search — the engine. What semantic search and RAG bring to cultural-heritage work and why (for researchers), plus the full architecture, pipeline, API and internals (for developers).
  • AI Assistant — the in-app chat surface that consumes the engine, and its own agent/MCP integration surface.

Pick your document

For researchers & curators — what and why

  • RAG & Semantic Search — Parts I–IV: what meaning-based search is, why it matters for heritage and memory, the use cases (cross-lingual search, "find records like this", grounded answers, dating an object from its image), and the limits & ethics. No programming required.
  • AI Assistant — overview — using the in-app chat to question your catalog in plain language.

For integrators — consume the API

  • RAG operational cookbook — the dd_rag_api actions (semantic_search, retrieve, ask, embed_groups, similar_objects, …), the in-client semantic search (quick input + panel block) with copy-paste request/response recipes.
  • Assistant cookbook — integrators — the assistant's SSE streaming protocol and HTTP recipes (dd_mcp_api).

For operators — install, connect, configure

For developers — architecture & internals

  • RAG & Semantic Search — Part V — the vector store, chunker, hybrid retrieval, ask() pipeline, the object-image stack, and the agent loop & MCP server. Code: src/ai/rag/, src/ai/agent/, src/ai/mcp/.

Two rules that never bend

  1. Opt-in everywhere. Nothing is indexed or answerable until you switch RAG on (DEDALO_RAG_ENABLED) and opt each section/component in via the ontology properties.rag. See the cookbook.
  2. ACL on every hit. Retrieval enforces the schema permission and the per-record projects filter before any result leaves the server — and the assistant/agent/MCP surfaces call straight into those same gated engines. The semantic layer can never become a back door around access control.

Related: SQO (precise structured search — the complement to semantic search) · Ontology (where opt-in is declared) · Configuration · Dédalo API v1.