RAG Setup Guide

Build a Retrieval-Augmented Generation (RAG) pipeline with DocLD to chat with your documents.

What is RAG?

RAG combines document retrieval with AI generation:

1. Retrieve - Find relevant document chunks
2. Augment - Add context to the AI prompt
3. Generate - Create answers using the context

This produces accurate, citation-backed responses from your documents.

Quick Start

1. Create a Knowledge Base

curl -X POST "/api/knowledge-bases" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Company Policies",
    "description": "HR policies and procedures"
  }'

2. Add Documents

Upload documents to your knowledge base:

curl -X POST "/api/upload" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F "file=@employee-handbook.pdf" \
  -F "knowledge_base_id=kb-uuid"

Or add existing documents:

curl -X POST "/api/knowledge-bases/{kb_id}/documents" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{"document_id": "doc-uuid"}'

3. Chat with Documents

The chat API is stream-only and uses the AI SDK UIMessage  contract. Create a session first, then send messages with message (UIMessage with role and parts), session_id, and knowledge_base_id. See the Chat API for full details.

Create a session:

curl -X POST "/api/chat/sessions" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"knowledge_base_id": "kb-uuid"}'

Send a message (use the session_id from the response):

curl -X POST "/api/chat" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "message": { "role": "user", "parts": [{ "type": "text", "text": "What is the vacation policy?" }] },
    "session_id": "session-uuid",
    "knowledge_base_id": "kb-uuid"
  }'

How RAG Works in DocLD

Document Processing Pipeline

Upload → Parse → Chunk → Vectorize → Index

1. Parse - Extract text, tables, figures from documents
2. Chunk - Split content into semantic units
3. Vectorize - Chunk text is sent to Pinecone; embeddings are generated server-side (llama-text-embed-v2)
4. Index - Records stored in Pinecone for semantic search

Query Pipeline

Question → Search (Pinecone embeds) → Retrieve → Generate → Respond

1. Search - Query text is sent to Pinecone; embeddings and similarity search happen server-side
2. Retrieve - Get top-k relevant chunks
3. Generate - AI creates answer using context
4. Respond - Return answer with citations

Configuring RAG

Retrieval Settings

Configure how many results to retrieve:

{
  "settings": {
    "retrieval": {
      "top_k": 5,
      "threshold": 0.7,
      "reranking": true
    }
  }
}

Chunking Strategy

Control how documents are split:

{
  "settings": {
    "chunking": {
      "strategy": "semantic",
      "max_size": 1000,
      "overlap": 100
    }
  }
}

Best Practices

Document Preparation

1. Quality over quantity - Curate relevant content
2. Consistent formatting - Well-structured documents work better
3. Remove noise - Exclude irrelevant sections
4. Update regularly - Keep content current

Knowledge Base Organization

1. Single topic - One domain per knowledge base
2. Appropriate size - Not too small, not too large
3. Related content - Documents should relate to each other

Query Optimization

1. Be specific - Clear questions get better answers
2. Context helps - Provide background when needed
3. Follow up - Use conversation context

Advanced Configuration

Hybrid Search

Combine vector search with keyword matching:

{
  "settings": {
    "retrieval": {
      "hybrid_search": true,
      "keyword_weight": 0.3
    }
  }
}

Response Modes

curl -X POST "/api/chat" \
  -d '{
    "message": "Explain the refund policy in detail",
    "knowledge_base_id": "kb-uuid",
    "mode": "thorough"
  }'

Evaluating RAG Quality

Metrics to Track

Improving Quality

1. Add more documents - Better coverage
2. Tune retrieval - Adjust top_k and threshold
3. Enable reranking - Improve relevance
4. Review feedback - Learn from user signals

Example: Legal Document RAG

Setup

# Create knowledge base for contracts
curl -X POST "/api/knowledge-bases" \
  -d '{"name": "Contract Library", "description": "Client contracts"}'
 
# Upload contracts
for file in contracts/*.pdf; do
  curl -X POST "/api/upload" \
    -F "file=@$file" \
    -F "knowledge_base_id=kb-uuid"
done

Query

curl -X POST "/api/chat" \
  -d '{
    "message": "What are the standard payment terms across our contracts?",
    "knowledge_base_id": "kb-uuid",
    "mode": "thorough"
  }'

Response

{
  "message": "Based on the contracts in your library, standard payment terms are Net 30 days. However, there are variations...",
  "citations": [
    {
      "text": "Payment shall be due within thirty (30) days...",
      "document_name": "Acme Contract.pdf",
      "page": 5
    }
  ],
  "confidence_score": 0.92
}

Troubleshooting

Low Confidence Scores

• Add more relevant documents
• Check document quality
• Adjust retrieval threshold

Irrelevant Results

• Increase relevance threshold
• Enable reranking
• Review chunking settings

Missing Information

• Ensure documents are fully processed
• Check if content is in the knowledge base
• Verify document parsing succeeded