Skip to main content

Overview

Intelligent document retrieval with grounded, citation-aware summarization

What is Document Search & Summarization?​

A production-ready service that combines:

  • Hybrid Search: BM25 + Vector embeddings for comprehensive retrieval
  • Grounded Summarization: Extractive and abstractive with citations
  • Profile-Based Architecture: Different SLOs for different use cases

Quick Comparison​

ProfileLatencyQualityCostUse Case
Balanced500msGood (0.7-0.8)$0.60/1KGeneral Q&A
Latency-First250msAcceptable (0.7)$0.35/1KInteractive
Quality-First5sExcellent (0.85-0.95)$52/1KResearch

Quick Start​

from packages.rag.document_search import DocumentSearchPipeline, ProfileType

# Initialize with profile
pipeline = DocumentSearchPipeline.create_profile(ProfileType.BALANCED, store, embed_fn)

# Execute
result = pipeline.execute("Your query")
print(result.summary.text)  # Grounded summary with citations

Full Example: Document Search Demo

Documentation Structure​

This guide is organized into focused topics:

  1. Overview (you are here) - Introduction and quick start
  2. Architecture - Profile-based design and system architecture
  3. Storage & Indexing - Document storage and indexing strategies
  4. API Integration - API endpoints and integration guide
  5. Complete Guide - Comprehensive documentation
  6. Quick Reference - Cheat sheet and quick start

Key Features​

Hybrid Retrieval​

  • BM25: Exact keyword matching (fast, explainable)
  • Vector Search: Semantic similarity (handles synonyms)
  • RRF Fusion: Combines both with reciprocal rank fusion

Grounded Summarization​

  • Extractive (TextRank): Fast, faithful, free
  • Abstractive (LLM): Fluent, comprehensive
  • Citations: Sentence-level citation tracking
  • Faithfulness: Verification with fail-closed design

Profile-Based Configuration​

  • Balanced: General-purpose (500ms, good quality)
  • Latency-First: Interactive speed (250ms)
  • Quality-First: Research-grade (5s, 95%+ faithfulness)

Implementation Files​

Core Module: packages/rag/document_search/

FilePurpose
pipeline.pyMain orchestrator with profiles
store.pyDocumentStore + OpenSearch
retriever.pyHybrid retrieval + query expansion
summarizer.pyGrounded summarization
test_fixtures.pyTest dataset (10 queries)

For implementation details, see the complete guide and architecture documentation.

Learning Paths​

Beginners (60 minutes)​

  1. This overview (10 min)
  2. Quick Reference (15 min)
  3. Run demo script (20 min)
  4. Architecture basics (15 min)

Practitioners (90 minutes)​

  1. Quick Reference (10 min)
  2. Architecture (30 min)
  3. Storage & Indexing (25 min)
  4. API Integration (25 min)

ML Engineers (2 hours)​

  1. Complete Guide (60 min)
  2. Architecture (30 min)
  3. Storage & Indexing (30 min)

Real-World Applications​

Customer Support KB​

Profile: Balanced
SLO: < 500ms, 85%+ faithfulness
Cost: $0.60 per 1K queries

result = pipeline.execute(
    "How do I reset my password?",
    filters={"category": "account"}
)

Legal/Compliance Research​

Profile: Quality-First
SLO: < 5s, 95%+ faithfulness
Cost: $52 per 1K queries

result = quality_pipeline.execute(
    "What is our data retention policy?",
    filters={"document_type": "policy", "approved": True}
)

Interactive Chat​

Profile: Latency-First
SLO: < 250ms, 70%+ relevancy
Cost: $0.35 per 1K queries

result = fast_pipeline.execute(partial_query)

Next Steps​

Quick Links: