document_search

Document Search

AI-powered document retrieval for banking operations, combining intelligent indexing with semantic search to help staff quickly locate policies, procedures, compliance documents, and regulations.

Overview

The Document Search application indexes banking documents by extracting metadata (type, category, tags, effective date) and then performs semantic search across the corpus to return ranked, relevant results with contextual snippets. It supports filtering by document type and status to ensure staff find current, authoritative documents.

Business Value

• Faster Document Retrieval -- Semantic search finds relevant documents in seconds versus manual browsing
• Improved Compliance -- Staff can quickly locate current policies and regulations when needed for audits or decisions
• Reduced Risk -- Active/archived/superseded status tracking prevents reliance on outdated documents
• Cross-Reference Discovery -- Automatic identification of related documents surfaces connections staff might miss
• Operational Efficiency -- Banking operations teams spend less time searching and more time acting

Architecture

graph TB
    Request["Client Request<br/><small>Document Search Query</small>"] --> Runtime["AgentCore Runtime"]
    Runtime --> Orchestrator["Orchestrator"]
    
    Orchestrator --> Indexer["Document Indexer<br/><small>Metadata extraction & categorization</small>"]
    Orchestrator --> SearchAgent["Search Agent<br/><small>Semantic search & ranking</small>"]
    
    Indexer --> Bedrock["Amazon Bedrock<br/>(Claude)"]
    SearchAgent --> Bedrock
    
    Indexer --> S3["S3 Sample Data<br/><small>data/samples/document_search/</small>"]
    SearchAgent --> S3
    
    Indexer --> Synthesis["Result Synthesis"]
    SearchAgent --> Synthesis
    
    Synthesis --> Response["Response<br/><small>Ranked results & snippets</small>"]
    
    style Request fill:#e1f5ff
    style Response fill:#d4edda
    style Bedrock fill:#fff3cd
    style S3 fill:#fff3cd

Loading

Directory Structure

use_cases/document_search/
├── README.md
└── src/
    ├── __init__.py                              # Framework router
    ├── strands/
    │   ├── __init__.py
    │   ├── config.py                            # Document search settings (document_ids list)
    │   ├── models.py                            # SearchRequest / SearchResponse
    │   ├── orchestrator.py                      # DocumentSearchOrchestrator
    │   └── agents/
    │       ├── document_indexer.py              # DocumentIndexer agent
    │       └── search_agent.py                  # SearchAgent agent
    └── langchain_langgraph/                     # LangGraph implementation (same structure)

Agentic Design

The DocumentSearchOrchestrator extends StrandsOrchestrator and implements a parallel fan-out pattern:

1. Parallel Indexing and Search -- The Document Indexer and Search Agent always run concurrently via asyncio.gather(), regardless of document type filter.
2. Document Type Filtering -- When a specific document_type is requested, the type filter is appended to the agent input text.
3. Config-Driven Document IDs -- Both agents read the list of document IDs from config.py settings and iterate over them with the s3_retriever_tool.
4. Synthesis -- A supervisor LLM call ranks and deduplicates results, producing a final summary with total documents found, relevance scores, and refinement suggestions.

Agents

Document Indexer

Field	Detail
Class	DocumentIndexer(StrandsAgent)
Role	Indexes and categorizes banking documents, extracting metadata and cross-references
Data	Document data via s3_retriever_tool (iterates over configured document IDs)
Produces	Document ID, title, type, category, tags, effective date, version, status, related documents, content summary

Search Agent

Field	Detail
Class	SearchAgent(StrandsAgent)
Role	Performs semantic search across the document corpus, returning ranked results
Data	Document data via s3_retriever_tool (iterates over configured document IDs)
Produces	Ranked document list with relevance scores, contextual snippets, metadata, search summary, related search suggestions

Data and Tools

• Tool: s3_retriever_tool -- Retrieves document data from S3 by document ID
• S3 Path: data/samples/document_search/{document_id}/
• Data Files: profile.json (document title, content, metadata, category)

Request / Response

Request (SearchRequest)

class SearchRequest(BaseModel):
    query: str                                     # e.g. "anti-money laundering requirements"
    document_type: DocumentType = "full"            # full | policy | procedure | compliance | regulation | guideline
    additional_context: str | None = None

Response (SearchResponse)

class SearchResponse(BaseModel):
    query: str
    search_id: str                                 # UUID
    timestamp: datetime
    results: list[SearchResult]                    # document_id, title, snippet, relevance, document_type, status
    relevance_scores: list[float]
    summary: str                                   # Executive summary of search results
    raw_analysis: dict

Document Statuses: active, archived, draft, superseded

Quick Start

# Deploy to AgentCore
USE_CASE_ID=document_search ./scripts/deploy/full/deploy_agentcore.sh

# Test
./scripts/use_cases/document_search/test/test_agentcore.sh

Sample Data

Document ID	Type	Description
DOC001	Policy	Anti-Money Laundering Policy

Related Documentation

• Platform Overview
• Architecture Patterns
• Deployment Guide