SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsai-knowledge-management-platform

ai-knowledge-management-platform

Data & analytics

View repository

Updated 10-29 04:35

About this skill

Agentic Documents Assistant (AWS sample)

Instructions

  1. Confirm this exact repository. It is an AWS-based example app, not a general-purpose knowledge-management platform. Pin source claims to references/source.json; inspect the user's checkout before relying on old commands.
  2. Trace the data path: ingestion notebooks/pipelines extract document text and structured entities; semantic retrieval uses an embedding model and an existing PGVector index; analytical questions use SQL over extracted tables; the LangChain agent can select retrieval, SQL, or tools. Keep these paths distinct in diagnosis.
  3. Use the concrete backend topology. Pinned CDK backend describes API Gateway, Lambda container/Bedrock orchestration, Cognito, DynamoDB chat history, Aurora PostgreSQL for semantic search and SQL, SSM parameters, VPC and SageMaker ingestion permissions. Do not assume a component is deployed or configured from source alone.
  4. Troubleshoot at the failing boundary: source preprocessing/load, vector index/embedding compatibility, retrieval settings, SQL schema/query, agent routing, API/auth, or AWS deployment. Verify the target dataset and infrastructure before any mutation.
  5. Gate every external effect separately. Approval is required before dependency/source installation, AWS login or credential/config changes, private-document access/import, cloud/billing/resource use, database reads/writes, CDK deploy or deletion, public service activation, or paid model/provider uploads. A request to inspect code authorizes none of these.
  6. Protect data and claims. Never request or reveal secrets. Do not execute generated SQL or deploy a stack as a troubleshooting shortcut. Distinguish source-defined behavior from a live AWS deployment or runtime result.
  7. Use references/upstream.md for setup/operations, source-specific failure modes and pinned URLs. The offline preflight helper hashes selected files only and does not verify AWS, dependencies, or runtime. Evals are authored scenarios, not run evidence.

Examples

  • "The assistant answers contract questions but misses expiry-date questions." → Trace whether extraction populated the relational entity table and whether SQL QA receives the expected schema; do not assume vector retrieval is responsible.
  • "Deploy this to our AWS account." → Explain the CDK path, then request separate approval for credentials, cloud resource creation/billing, and public exposure.
  • "Compare knowledge-management products for my team." → Route to product/market research rather than this sample-specific skill.

Best practices

  • Treat the sample's infrastructure as an example architecture; validate account, region, quotas, costs and current AWS APIs before any deployment plan.
  • Keep vector retrieval (PGVector existing index) separate from SQL analytics (structured extracted table); verify ingestion and schema first.
  • Follow least privilege and scope source documents, tables, and cloud permissions explicitly.
  • A successful local build, CDK synth, or offline source check is not proof of deployed or secure runtime behavior.

References