innovatieplatform/wiki/concepts/ai-integration.md

title, created, updated, status, tags, sources

title	created	updated	status	tags	sources
AI Integration	2026-04-08	2026-04-08	speculative	concept ai rag langgraph embeddings	ai-service/app/main.py ai-service/requirements.txt docker-compose.yml

AI Integration

Current State

The AI service is a Python FastAPI stub with placeholder endpoints. No actual AI processing is wired up yet.

Implemented (stub only)

Endpoint	Method	Status
GET /health	Health check	Working
POST /api/chat	Chat with context	Stub — returns placeholder text
POST /api/summarize	Generate summaries	Stub — returns placeholder text
POST /api/search	Semantic search	Stub — returns empty results

Request/Response Models (Pydantic)

ChatRequest:     message, project_id?, conversation_history[]
ChatResponse:    reply, project_id?
SummarizeRequest: content, project_id?, summary_type?
SummarizeResponse: summary, project_id?
SearchRequest:   query, project_id?, limit?
SearchResponse:  results[{id, content, score, metadata}], query

Planned Architecture

Laravel App ↔ HTTP ↔ Python AI-Service (FastAPI)
                        ├── LangGraph Orchestrator
                        │     ├── Router / Classifier
                        │     └── Agent graph (state machine)
                        ├── Anthropic Claude (LLM)
                        ├── pgvector (embeddings / similarity search)
                        └── Tools:
                              ├── DB query (project data, commitments, phases)
                              ├── Document retrieval (semantic search)
                              └── Embedding generation

RAG Pipeline (planned)

Sources

• Project descriptions and phase notes
• Documents (uploaded files, meeting notes)
• Lessons learned
• Decisions and their rationale
• Knowledge articles

Embedding Strategy

• Storage: pgvector extension on PostgreSQL 16
• Models: Document and KennisArtikel already have embedding vector columns
• Update triggers: On document create/update, on project phase change
• Chunking: Per document type and size

Agent Skills (from CLAUDE.md)

Agent	Autonomy	Purpose
Project Assistant	Low	Answer questions about specific projects
Knowledge Assistant	Low	Search and surface knowledge articles
Document Assistant	Medium	Summarize, compare, extract from documents
System Tasks	High	Background indexing, embedding updates

Content Governance Rules

1. AI-generated content always labeled ("AI-suggestie", "Concept")
2. Human confirmation required before AI content gains system status
3. All AI interactions logged (request, response, tools used, sources cited)
4. Source attribution mandatory in AI responses
5. Confidence indicators when certainty is low