2.3 KiB
2.3 KiB
Step 2 Flat File Context Design (Website Analysis)
Intent
Step 2 context must be optimized for AI-agent retrieval speed and token efficiency, not human readability.
Current storage location
workspace/workspace_<safe_user_id>/agent_context/step2_website_analysis.json
Current retrieval chain
- Flat file (fastest)
- DB (
website_analyses) - SIF semantic fallback
Compactness strategy
For implementation, keep two logical layers:
dequivalent (full canonical data) for deep reasoning.sequivalent (high-signal summary) for fast agent prompts and most decisions.document_contextfor machine-readable orientation (purpose, journey stage, fallback contract, context-window guidance).
Agents should default to summary-first reads and only open full data when needed.
Step 2 coverage requirements
The Step 2 context should preserve these semantic groups:
- identity/state: website url, timestamps, status/error/warning
- brand/style: writing style, style patterns/guidelines, brand analysis
- audience/content: target audience, content type, recommended settings, characteristics
- strategy/seo: strategy insights, SEO audit, strategic history
- crawl/discovery: crawl output, meta info, sitemap analysis
- traceability: raw inbound payload snapshots
Agent-readability best practices
- Keep keys stable and deterministic.
- Prefer arrays/enums over long free text.
- Keep summary fields flattened and high signal.
- Avoid duplicate verbose nested structures unless required for correctness.
- Include retrieval hints for consistent downstream querying.
Practical guidance for consumers
- Use summary/high-signal fields first for routing and lightweight reasoning.
- Pull deep fields only for specialist tasks (SEO, persona fidelity, editorial style checks).
- If flat-file missing/stale: auto-fallback to DB then SIF.
Note
A generalized compact framework is documented in:
docs/flat_file_context/FLAT_FILE_CONTEXT_FRAMEWORK_DESIGN.md
Future enhancements are tracked in:
docs/flat_file_context/FLAT_FILE_CONTEXT_ENHANCEMENTS_BACKLOG.md
Context window guidance
- Keep summary compact and deterministic.
- Add byte-size metadata to help agents decide whether to expand into full data.
- Prefer short keys and avoid verbose natural language in machine envelopes.