Skip to content

Latest commit

 

History

History
192 lines (143 loc) · 10.1 KB

knowledge-base.md

File metadata and controls

192 lines (143 loc) · 10.1 KB
title Document Knowledge Base Query

Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved. SPDX-License-Identifier: MIT-0

Document Knowledge Base Query

The GenAIIDP solution includes an integrated Document Knowledge Base query feature that enables you to interactively ask questions about your processed document collection using natural language. This feature leverages the processed data to create a searchable knowledge base.

S3Vectors.mp4

How It Works

  1. Document Processing & Indexing

    • Processed documents are automatically indexed in a vector database
    • Documents are split into fixed-size chunks of up to 8,192 tokens, with roughly 15% overlap between consecutive chunks. The main template selects fixed-size chunking and the token limit, and the overlap is the knowledge-base nested stack's default; none of these is a stack parameter, and the chunking strategy cannot be changed after the knowledge base has been created
    • Each chunk maintains reference to its source document
    • Ingestion Schedule: Documents are ingested into the knowledge base every 30 minutes, so newly processed documents may not be immediately available for querying
  2. Interactive Query Interface

    • Access through the Web UI via the "Knowledge Base" section
    • Ask natural language questions about your document collection
    • View responses with citations to source documents
    • Follow-up with contextual questions in a chat-like interface
  3. AI-Powered Responses

    • LLM generates responses based on relevant document chunks
    • Responses include citations to source documents
    • Links to original documents for reference
    • Context-aware for follow-up questions

Query Features

  • Natural Language Understanding: Ask questions in plain English rather than using keywords or query syntax
  • Document Citations: Responses include references to the specific documents used to generate answers
  • Contextual Follow-ups: Ask follow-up questions without repeating context
  • Direct Document Links: Click on document references to view the original source
  • Markdown Formatting: Responses support rich formatting for better readability
  • Real-time Processing: Get answers in seconds, even across large document collections

Architecture & Vector Storage Options

The Knowledge Base feature supports two vector storage backends to optimize for different performance and cost requirements:

Vector Store Comparison

Aspect OpenSearch Serverless S3 Vectors
Query Latency Sub-millisecond Sub-second
Pricing Model Always On (continuous capacity costs) On Demand (pay-per-query)
Storage Cost Higher 40-60% lower
Best For Real-time applications Cost-sensitive deployments
Features Full-text search, advanced filtering Native S3 integration

Choosing Your Vector Store

  • S3 Vectors (Default): Choose for cost optimization with acceptable sub-second query latency
  • OpenSearch Serverless: Choose for applications requiring ultra-fast retrieval and real-time performance

Configuration

The Document Knowledge Base Query feature is configured with these stack parameters at deployment time:

Parameter Values Purpose
DocumentKnowledgeBase DISABLED, BEDROCK_KNOWLEDGE_BASE (Create) (default) Turns the feature on. The "Create" value is what provisions the knowledge base and its vector store.
KnowledgeBaseVectorStore S3_VECTORS (default), OPENSEARCH_SERVERLESS Which vector store backs the knowledge base. See the comparison above.
KnowledgeBaseEmbeddingModel amazon.titan-embed-text-v2:0 (default), cohere.embed-english-v3, cohere.embed-multilingual-v3 The Bedrock model that embeds documents into the vector store. See Embedding Model.
KnowledgeBaseModelId A Bedrock model id, default us.amazon.nova-pro-v1:0 The model that generates the answer from the retrieved passages.

KnowledgeBaseModelId has an explicit AllowedValues list, so the picker in the CloudFormation console shows exactly what is selectable; read it from template.yaml rather than from a copy here, because the list moves as models are added and retired. Two region-specific behaviours are worth knowing before you pick one:

  • Canada. A ca-central-1 or ca-west-1 stack left at the default is switched to ca.amazon.nova-lite-v1:0 automatically, because the default has no inference profile in Canada. An explicit choice is passed through unchanged and must be callable from the stack's region.
  • GovCloud. None of the us. / eu. / global. cross-region inference-profile prefixes resolve in GovCloud. Use one of the bare or us-gov. ids the parameter lists; the --govcloud template transform overrides the commercial default for you.

Whichever model you choose must have model access granted in Amazon Bedrock in the deployment region.

Note: ShouldUseDocumentKnowledgeBase is not a parameter you set. It is a CloudFormation Condition, derived from DocumentKnowledgeBase, which the parent stack passes down to the pattern stack as a boolean. Setting DocumentKnowledgeBase is the whole of the control.

Embedding Model

KnowledgeBaseEmbeddingModel selects the model that turns documents into vectors, for either vector store type. It accepts amazon.titan-embed-text-v2:0 (the default), cohere.embed-english-v3 and cohere.embed-multilingual-v3. All three produce 1024-dimension vectors, which is the size the vector index is created with.

Choose it when the knowledge base is first created. Changing it on an existing stack is not supported. With OpenSearch Serverless the update rebuilds the vector index, discarding the vectors already in it. With S3 Vectors the knowledge base keeps the embedding model it was created with, so the new choice does not take effect, and the vector index and its vectors are kept. The S3 Vectors index is rebuilt, which discards its vectors, only when an update changes its dimension, distance metric or metadata configuration, and the template sets all three to fixed values, so a stack parameter change does not rebuild it.

Deleting the stack deletes the S3 Vectors bucket and every index in it, including an index left under an earlier index name. S3 Vectors will not delete a bucket that still holds an index, so the custom resource lists the bucket's indexes and deletes each one first. This applies to the bucket the stack names for itself. When the nested stack's pS3VectorBucketName parameter names the bucket, which may be one that existed before the stack, only the stack's own index is deleted, and a bucket that still holds other indexes is left in place. In either case, if the bucket cannot be deleted, the stack delete fails with a message naming the bucket rather than reporting success. An update that changes only the index name creates the new index and keeps the old one. For the bucket the stack names for itself, the stack delete then removes the old index; for a bucket named by pS3VectorBucketName, the old index stays and the stack delete fails until it is deleted.

Grant model access for the model you choose in Amazon Bedrock, in the deployment region, before deploying with this feature enabled — without it, ingestion fails. When GovCloud model availability was checked on 2026-07-17, amazon.titan-embed-text-v2:0 was the only embedding model offered there, so keep the default in GovCloud unless you have confirmed the alternative is available.

When the feature is enabled, the solution:

  • Creates the selected vector storage resources (OpenSearch or S3 Vectors)
  • Configures API endpoints for querying the knowledge base
  • Adds the query interface to the Web UI

Using the Knowledge Base

Accessing the Knowledge Base

  1. Log in to the Web UI
  2. Navigate to the "Knowledge Base" section in the main navigation
  3. You'll see a chat-like interface for querying your document collection

Knowledge Base Query Tool answering a question about a bank statement, with its source document cited

Asking Questions

  1. Type your question in the input field at the bottom of the screen
  2. Press Enter or click Submit
  3. The system will process your question and return an answer
  4. The answer will include:
    • A direct response to your question
    • Citations to the source documents
    • Links to view the original documents

Exploring Document Context

  1. Click on document citations to view the original source
  2. The system will highlight the relevant sections in the document
  3. You can navigate to other parts of the document to explore the full context

Follow-up Questions

  1. After receiving an answer, you can ask related follow-up questions
  2. The system maintains context from previous questions
  3. This allows for a natural conversation about your documents
  4. You can start a new topic at any time by asking an unrelated question

Best Practices

  1. Be specific: Clearly state what information you're looking for
  2. Start broad, then narrow: Begin with general questions before diving into specifics
  3. Use follow-ups: Build on previous questions to explore topics in depth
  4. Check citations: Verify information by consulting the source documents
  5. Refine questions: If you don't get the expected answer, try rephrasing your question

Performance Considerations

  • Document Collection Size: Performance may vary with very large document collections
  • Query Complexity: More complex queries may take longer to process
  • Document Types: Some document types may be indexed more effectively than others
  • Model Selection: Different Bedrock models offer different performance/accuracy tradeoffs

Security Considerations

The Knowledge Base feature maintains the security controls of the overall solution:

  • Access is restricted to authenticated users
  • Document visibility respects user permissions
  • Questions and answers are processed securely within your AWS account
  • No data is sent to external services beyond the configured Bedrock models