Document Search & Summarization - Quick Reference

One-page cheat sheet for quick lookup

---

Three Profiles at a Glance

Profile	Latency	Quality	Cost	When to Use
Balanced	500ms	0.7-0.8	$0.60/1K	General Q&A, customer support
Latency-First	250ms	0.7	$0.35/1K	Interactive search, auto-complete
Quality-First	5s	0.85-0.95	$52/1K	Research, compliance, legal

---

Quick Start (3 Steps)

# 1. Setup
from packages.rag.document_search import *

store = OpenSearchDocumentStore("localhost", 9200, "documents")
pipeline = DocumentSearchPipeline.create_profile(
    ProfileType.BALANCED, store, embedding_function
)

# 2. Execute
result = pipeline.execute("Your query here")

# 3. Use Results
print(result.summary.text)  # Summary
print(result.summary.citations)  # Citations
print(result.timing)  # Performance metrics

---

Architecture

Query → Retriever → Reranker → Summarizer → Result
         ↓            ↓           ↓
       (Hybrid)   (Optional)  (Grounded)
       BM25+Vec   CrossEnc    Extract/Abstract

---

Key Concepts

Hybrid Search = BM25 + Vector

• BM25: Exact keyword matching (fast, explainable)
• Vector: Semantic similarity (handles synonyms)
• RRF: Combines both with reciprocal rank fusion
• α: Weight parameter (0=BM25 only, 1=Vector only, 0.5=equal)

Query Expansion

• PRF (Pseudo-Relevance Feedback): Use top results to expand query
• HyDE (Hypothetical Doc Embeddings): Generate ideal doc, search for similar

Reranking

• Bi-Encoder (retrieval): Fast, approximate similarity
• Cross-Encoder (reranking): Slow, accurate relevance scoring
• Two-Stage: Bi-encoder gets 50 candidates → Cross-encoder ranks → Top 5

Summarization

• Extractive: Select important sentences (fast, faithful, free)
• Abstractive: Generate new text (fluent, expensive, risk of hallucination)
• Grounded: Verify claims against sources, cite everything

---

Configuration Patterns

Pattern 1: Customer Support KB

config = PipelineConfig(
    profile=ProfileType.BALANCED,
    topK=20,
    alpha=0.5,
    enable_reranking=True,
    summarization_mode=SummarizationMode.EXTRACTIVE,
    max_summary_length=250
)

Pattern 2: Legal/Compliance

config = PipelineConfig(
    profile=ProfileType.QUALITY_FIRST,
    topK=50,
    alpha=0.5,
    enable_reranking=True,
    query_expansion_method="hyde",
    summarization_mode=SummarizationMode.ABSTRACTIVE,
    faithfulness_threshold=0.95
)

Pattern 3: Real-Time Chat

config = PipelineConfig(
    profile=ProfileType.LATENCY_FIRST,
    topK=10,
    alpha=0.7,  # Favor BM25 for speed
    enable_reranking=False,
    enable_query_expansion=False,
    summarization_mode=SummarizationMode.EXTRACTIVE,
    max_summary_length=150
)

---

RAGAS Metrics

Metric	Formula	Target (Balanced)	Target (Quality)
Context Precision	Relevant/Retrieved	> 0.70	> 0.85
Context Recall	Retrieved/Total Relevant	> 0.70	> 0.85
Faithfulness	Supported Claims/Total Claims	> 0.85	> 0.95
Answer Relevancy	cosine(query, answer)	> 0.75	> 0.80

---

Common Tuning Parameters

Retrieval

topK = 20  # Number of results to retrieve
alpha = 0.5  # BM25 vs vector weight (0-1)

BM25

k1 = 1.2  # Term frequency saturation (0.5-3.0)
b = 0.75  # Length normalization (0-1)

RRF

k = 60  # Rank constant (typically 60)

Reranking

model = "cross-encoder/ms-marco-MiniLM-L-6-v2"  # Balanced
model = "cross-encoder/ms-marco-MiniLM-L-12-v2"  # Quality
top_k = 10  # Rerank top N candidates

Summarization

max_length = 250  # Words
faithfulness_threshold = 0.85  # Minimum acceptance

---

Caching Strategy

# L1: Full result cache (1 hour TTL)
cache_key = f"query:{hash(query)}:{hash(filters)}"

# L2: Retrieval cache (1 hour TTL)
cache_key = f"retrieval:{hash(query)}"

# L3: Summary cache (24 hour TTL)
cache_key = f"summary:{doc_ids_hash}"

# L4: Embedding cache (7 day TTL)
cache_key = f"embedding:{content_hash}"

Expected Savings: 40-60% cost reduction

---

Error Handling

try:
    result = pipeline.execute(query)
except TimeoutError:
    # Fall back to faster profile
    result = latency_pipeline.execute(query)
except Exception as e:
    # Log and return error
    log_error(f"Pipeline failed: {e}")
    return ErrorResponse("Service unavailable")

---

Monitoring Metrics

# Latency
emit("search.latency_ms", timing["total_ms"])

# Quality
emit("search.faithfulness", summary.faithfulness)
emit("search.slo_met", slo_met)

# Usage
emit("search.queries_total", 1)
emit("search.cache_hit", cache_hit)

# Cost
emit("search.cost_usd", cost)

---

API Quick Reference

# Create pipeline
pipeline = DocumentSearchPipeline.create_profile(profile, store, embed_fn)

# Execute search
result = pipeline.execute(query, filters={...})

# Access results
result.query           # Original query
result.results         # List[RetrievalResult]
result.summary         # GroundedSummary
result.timing          # Dict[str, float]
result.slo_met         # bool
result.facets          # Dict (if enabled)

# Summary details
result.summary.text              # Summary text
result.summary.citations         # Dict[int, Citation]
result.summary.faithfulness      # float
result.summary.coverage          # float
result.summary.key_points        # List[str]

# Citation details
citation = result.summary.citations[1]
citation.document_id             # str
citation.document_title          # str
citation.chunk_id                # str
citation.snippet                 # str (100-150 chars)

---

Troubleshooting

Issue	Likely Cause	Solution
Slow searches	Profile misconfigured	Use latency-first, enable caching
Low faithfulness	Poor source quality	Use extractive mode, verify sources
Poor relevance	Wrong alpha weight	Tune α, enable query expansion
High costs	Too many LLM calls	Use extractive, implement caching
SLO violations	Too ambitious targets	Adjust profile or hardware

---

Performance Targets

Profile	Latency P95	Throughput	Concurrent Users
Balanced	< 500ms	100 QPS	1000+
Latency-First	< 250ms	200 QPS	2000+
Quality-First	< 5000ms	20 QPS	200+

---

Cost Breakdown (per 1K queries)

Component	Balanced	Latency-First	Quality-First
Retrieval	$0.40	$0.25	$1.50
Reranking	$0.20	$0.00	$0.50
Summarization	$0.00	$0.00	$50.00
Embedding	$0.10	$0.10	$0.50
Total	$0.60	$0.35	$52.50

---

Decision Tree

Start
  ↓
Need < 250ms latency? → Yes → Latency-First
  ↓ No
Need 95%+ faithfulness? → Yes → Quality-First
  ↓ No
General use case? → Yes → Balanced
  ↓ No
Custom requirements? → Create custom profile

---

Further Reading

• Full Guide: Document Search & Summarization
• Examples: Document Search Examples
• API Reference: Document Search API
• Theory: Information Retrieval section in main guide

---

Last Updated: October 9, 2025
Version: 1.0