Added Citation System To My n8n RAG Agent (Now Tells Me Which Document Every Answer Came From) 🔥

RAG agent gave great answers. Client asked: 'Which contract did that come from?' Agent couldn't say. Added citation tracking. Now every answer includes source document and page number.

THE PROBLEM WITH BASIC RAG:

Agent answers questions. Doesn't track source documents. Can't verify answers. Can't cite sources. Legal teams need sources for every claim.

THE SOLUTION:

Build citation tracking into RAG pipeline. Capture document metadata. Pass through embeddings. Return with answers.

THE EXTENSION:

ORIGINAL RAG FLOW:

Document → Parse → Split → Embed → Store → Query → Answer

EXTENDED RAG FLOW:

Document → Parse → Add Metadata → Split (preserve metadata) → Embed (with metadata) → Store with source → Query → Retrieve with source → Answer with citation

THE NEW METADATA STRUCTURE:

document_id, document_name, document_type, page_number, section, last_updated, owner

STORED WITH EACH CHUNK:

Every text chunk in vector store includes metadata. When agent retrieves chunks, metadata comes with it.

AGENT TOOL ENHANCEMENT:

Modified 'Search Documents' tool to return text PLUS source information.

Agent receives: Relevant text, source document name, page numbers, confidence score.

Agent includes in answer: 'According to the Service Agreement (page 7), the cancellation policy is...'

THE IMPLEMENTATION:

NODE CHANGES:

1. After Document Parser: SET node creates metadata object

2. Text Splitter: Configure to preserve metadata in chunks

3. Vector Store Insert: Include metadata with embeddings

4. Vector Store Query: Return metadata with results

5. Agent Tool: Format metadata into citation

CITATION FORMAT OPTIONS:

Simple: (Source: filename.pdf)

Detailed: (Source: filename.pdf, page 7, section 3.2)

Legal: (Service Agreement v2.1, executed 2024-01-15, section 3.2, page 7)

THE RESULTS:

- Base RAG: Answers without sources

- Extended RAG: Every answer cited

- Legal team: Approved (had rejected basic RAG)

- Audit compliance: 100% traceable

- User trust: Significantly increased

REAL EXAMPLE:

Question: 'What is our refund policy for enterprise customers?'

Basic RAG: 'Enterprise customers can request refunds within 30 days with manager approval.'

Citation-enabled RAG: 'Enterprise customers can request refunds within 30 days with manager approval. (Source: Enterprise Terms of Service v3.2, page 12, section 4.3, last updated 2024-03-01)'

Can verify. Can audit. Can trust.

CONFIGURATION:

Metadata schema: Define in SET node. Consistent across documents.

Vector store: Qdrant metadata filtering. Query specific document types or date ranges.

THE LESSON:

RAG without citations is demo. RAG with citations is production. Legal and compliance require source tracking.

TEMPLATE:

Complete citation-enabled RAG system. Metadata schema, vector store config, agent tool with citations.

How do you handle source tracking in RAG?

18 comments