README.md

AegisAI synthetic knowledge-base fixtures

These 30 files are fictional, safe-to-commit documents for exercising the full local pipeline: upload, background validation, extraction, chunking, OpenAI embeddings, Qdrant retrieval, streaming RAG chat, and later document-level access control. They contain no real credentials, customer data, or production policy.

The set intentionally overlaps. For example, search for refresh token, incident severity, document retention, SAML, outbox, or vector cleanup and compare retrieval and cited chat responses across several source documents.

Regenerate the PDF and DOCX fixtures when needed (the committed generator makes their contents reviewable):

backend/venv/bin/python sample-data/generate_enterprise_fixtures.py

Upload all files after obtaining a documents:write token:

for document in sample-data/knowledge-base/*.{md,txt,pdf,docx}; do
  [ -f "$document" ] && [ "$(basename "$document")" != 'README.md' ] || continue
  case "$document" in
    *.md) upload_type='text/markdown' ;;
    *.txt) upload_type='text/plain' ;;
    *.pdf) upload_type='application/pdf' ;;
    *.docx) upload_type='application/vnd.openxmlformats-officedocument.wordprocessingml.document' ;;
  esac
  curl --fail-with-body -X POST http://localhost:8000/documents \
    -H "Authorization: Bearer YOUR_ACCESS_TOKEN" \
    -F "file=@${document};type=${upload_type}"
done

Some curl installations label .md files as application/octet-stream by default. The explicit type=text/markdown is therefore required; AegisAI correctly rejects that generic binary MIME type.

Use GET /documents/{id}/indexing-status to wait for each document to reach succeeded. Then use the same user (or another user with documents:read) to test /retrieval/search and /chat/stream.

Enterprise demo scenarios

After all files are indexed, use these questions to test overlapping retrieval and grounded citations:

Scenario	Query	Expected source themes
Identity and access	How are local roles reviewed and why are SSO claims insufficient?	Access overview, SSO linking, quarterly access review, architecture review.
Incident response	What is required in the first fifteen minutes of a Severity 1 incident?	Incident severity, communications, incident-response playbook.
Secure data handling	Which data must never appear in a knowledge-base document?	Data classification, retention, privacy assessment.
Release engineering	What should be verified before and after a schema change release?	Change management, release readiness, onboarding.
Resilience	How should derived vectors be restored after a disaster?	Business continuity, disaster recovery, vector cleanup.
Grounded chat	When should the assistant refuse to answer from model knowledge?	RAG chat, customer support runbook, chat-history boundaries.

Today, every user with documents:read can retrieve this shared demo data. Phase 12 will use the same pack to demonstrate document grants, revocation, and different search/chat results for different enterprise users.

The fixture set contains 10 Markdown records, 10 TXT records, 5 PDFs, and 5 DOCX files. The source generator uses fictional content only; no credentials, real users, customer records, or production policies belong in this directory.