Amazon Kendra | Overview.

Amazon Kendra - Overview.

Scope:

• Quick elevator (one-liner),
• Core concepts & components,
• How Kendra “thinks” about relevance,
• Query types & behavior,
• GenAI integration (what’s different),
• Security & governance,
• Scalability, performance & capacity,
• Observability & metrics to watch,
• Costs & pricing model (high level),
• Integrations & common architectures,
• Best practices & tuning checklist,
• Common pitfalls & gotchas,
• Sample reference architecture (textual diagram),
• An Architecture diagram,
• Link to AWS Documentation,
• Insights.

Quick elevator (one-liner)

• Amazon Kendra is a managed, ML-powered enterprise search service that provides semantic / natural-language search over heterogeneous content (S3, SharePoint, databases, FAQs, etc.).
• Amazon Kendra is equipped with features for:
• Relevance tuning, 
• Connectors, 
• Document enrichment, 
• (newer) GenAI-index integration for retrieval-augmented responses.

Core concepts & components

• Index — the primary unit: stores documents, attributes, embeddings/representations, and ranking metadata. 
•  Index types include Developer, Enterprise, and GenAI Enterprise (for easier reuse with generative tools).
• Data sources / Connectors — native/scheduled connectors for S3, SharePoint, Salesforce, RDS, Confluence, Drive, etc. 
• twtech can also implement custom connectors or push via API.
• Ingestion pipeline — fetch (connector) → convert/parse (file types: PDF, Office, HTML, text, etc.) → optional pre-processing/enrichment → indexing. Custom Document Enrichment lets twtech alter attributes/content at ingest time.
• Document chunking & fields — documents are broken into passages/paragraph chunks; fields/metadata (author, source, date, doc-type) are indexed separately so twtech can boost/filter by them.
• Query engine — combines semantic matching (deep models / embeddings), lexical matching, and a ranking model that uses field boosts, recency, and other signals to produce a ranked result set.
• Supports direct answers (answer extraction) and returning source passages.

How Kendra “thinks” about relevance

• Hybrid scoring — Kendra blends semantic similarity (contextual/embedding-like signals) with traditional retrieval and engineered ranking features.
• Relevance tuning — twtech can boost documents by field, source, or attribute (e.g., boost FAQs, fresh docs, or a trusted source). This adjusts which documents are favored in ranking; it doesn’t force inclusion but biases ranking. There are UI and API ways to tune.
• Synonyms & domain vocab — upload synonym lists and domain-specific terminology so queries with alternate terms still match the correct content.

Query types & behavior

• Natural-language questions (Who, What, How) → Kendra extracts concise answers when the exact answer is present in content.
• Descriptive answers → returns passages or full documents for how-to or explanatory queries.
• Keyword / faceted search → supports filters/facets on metadata and structured attributes.
• Limits — Kendra is not designed to aggregate answers requiring cross-document synthesis or to perform arbitrary calculations over content (it favors direct extraction and ranking).

GenAI integration (what’s different)

• GenAI Enterprise Index: allows Kendra-indexed data to be used as a retriever for generative systems (e.g., Bedrock, other RAG setups) — easier reuse of indexed knowledge for assistants and prompt flows.
• It also has more granular capacity options.

Security & governance

• Encryption — data encrypted at rest (AWS-owned/managed/customer KMS) and in transit (TLS).
• Access control — integrates with IAM and supports document-level access control (ACL propagation from source connectors) so search results can be filtered by user permissions.
• Audit / compliance — use CloudTrail and CloudWatch for logging queries, index actions, and connector activity.

Scalability, performance & capacity

• Capacity model — twtech provisions index capacity units (storage + query units) and these are billed per-hour for storage/query capacity; query throughput and latency correlate with provisioned capacity.
• Sync modes — connectors can schedule incremental syncs; for heavy-change environments twtech can push deltas via API to avoid full crawls.

Observability & metrics to watch

• Query latency,
• Query units utilization,
• Error rate for connectors,
• Document ingestion throughput,
• Freshness lag,
• Top-queries with no matches (helpful for gap analysis).
• Use CloudWatch dashboards + logs to track and alert.

Costs & pricing model (high level)

Main bill drivers:

• Number/type of indexes (index hours),
• Storage units for index content,
• Query units (query-hours and/or per-query costs depending on edition),
• Connectors and connector hours.
• GenAI index tiers may have different starting prices and finer-grained capacity.
• Always validate current pricing in twtech AWS account/region.

Integrations & common architectures

• Intranet / knowledge portal: Kendra index + Experience Builder (or custom UI) + authentication via your SSO + connectors to S3/SharePoint/Drive.
• Customer support / help center: Kendra behind a chat UI (Lex/third-party) to power FAQ answering and surfacing KB articles; use Custom Document Enrichment to surface product metadata/intent.
• RAG / virtual assistant: Kendra GenAI index used as retriever in a RAG pipeline (retriever → reranker → generator) to ground generative responses in company documents.

Best practices & tuning checklist

1. Start small, iterate: index a representative subset (top 5–10 content sources) and tune relevance there.
2. Enrich on ingest: normalize metadata (product IDs, region, department) via CDE(custom document enrichment) so twtech can filter/boost reliably.
3. Use synonyms & FAQs: capture domain synonyms and explicit FAQ entries for high-value Q→A mapping.
4. Relevance tuning: add source/freshness boosts for authoritative or time-sensitive content; measure impact with A/B tests (query logs).
5. Permissions: propagate ACLs from sources or enforce via attributes to prevent oversharing.
6. Instrument queries: capture zero-result queries and answer-quality feedback so twtech can iteratively re-index or add clarifying content.

Common pitfalls & gotchas

• Expectations mismatch: Kendra extracts direct answers only when the exact answer text exists; generative paraphrasing is limited unless twtech pairs with a generation model.
• Poor metadata: if sources lack structured attributes, tuning and faceting are weaker — invest in enrichment.
• Overfitting boosts: overly aggressive boosts can hide useful documents; monitor search quality after each change.

Sample reference architecture (textual diagram)

1. Sources: SharePoint, S3 (documents), RDS (DB), Salesforce. 
2. Connectors: scheduled syncs + webhook pushes for near real-time. 
3. Preprocessing: optional Lambda jobs / Comprehend / Transcribe to extract entities, OCR, or audio transcripts. 
4. Kendra Index (GenAI/Enterprise) with Custom Document Enrichment. 
5. Application layer: Search API / Experience Builder / Chat UI (with RAG: Kendra as retriever + generator service). 
6. Security & Ops: IAM, KMS, CloudWatch, CloudTrail.

An Architecture diagram image showing:

 Connectors → Kendra → App → Security.

Link to AWS Documentation: 

https://docs.aws.amazon.com/kendra/latest/dg/hiw-index-types.html?utm_source=chatgpt.com

twtech-Insights: