Skip to content

RAG Deployment

Deploy retrieval-augmented generation systems.

Quick Start

# 1. Ingest documents
docs2synth rag ingest --processed-dir data/processed/

# 2. Query
docs2synth rag run -q "What is the invoice total?"

# 3. Launch demo
docs2synth rag app

Strategies

Naive

Single retrieval + generation pass.

docs2synth rag run -s naive -q "Your question"

How it works: 1. Retrieve top-k documents 2. Generate answer with LLM using retrieved context

Iterative

Multi-step refinement with feedback.

docs2synth rag run -s iterative -q "Your question" --show-iterations

How it works: 1. Retrieve documents 2. Generate initial answer 3. Compare with previous iteration 4. Refine until convergence or max iterations

Document Ingestion

Index processed documents into vector store.

# Basic
docs2synth rag ingest

# With options
docs2synth rag ingest \
  --processed-dir data/processed/ \
  --processor docling \
  --include-context

What's indexed: - Each text object embedded separately - Metadata: source file, object_id, bbox, page, label - Context stored in metadata (not embedded)

Querying

CLI Query

# Basic query
docs2synth rag run -q "What is the total amount?"

# With strategy
docs2synth rag run -s iterative -q "What is the invoice date?"

# Show iteration details
docs2synth rag run -s iterative -q "Question" --show-iterations

Demo App

Interactive Streamlit interface:

docs2synth rag app
docs2synth rag app --host localhost --port 8501 --no-browser

Opens at http://localhost:8501

Features: - Interactive query interface - Document preview with retrieved chunks - Strategy comparison - Source attribution with confidence scores

Configuration

config.yml: 

rag:
  vector_store:
    type: chroma
    persist_directory: ./data/rag/vector_store
    collection_name: docs2synth_collection

  embedding:
    model: sentence-transformers/all-MiniLM-L6-v2

  strategies:
    naive:
      retriever:
        k: 5  # Top-k documents
      generator:
        provider: openai
        model: gpt-4o-mini
        temperature: 0.7

    iterative:
      retriever:
        k: 5
      generator:
        provider: openai
        model: gpt-4o-mini
        temperature: 0.7
      max_iterations: 3
      similarity_threshold: 0.9

Vector Store Management

# List strategies
docs2synth rag strategies

# Reset vector store (clears all documents)
docs2synth rag reset

Custom Strategies

Add custom RAG strategies by extending RAGStrategy class:

from docs2synth.rag.strategy import RAGStrategy

class MyCustomStrategy(RAGStrategy):
    def run(self, query: str, state: RAGState) -> RAGResult:
        # Implement custom retrieval + generation logic
        pass

Register in config: 

rag:
  strategies:
    custom:
      class: mymodule.MyCustomStrategy
      retriever:
        k: 10

Next Steps