5.3 KiB
5.3 KiB
Autofill: Learning, Personalization, and Explainability
This document outlines next-step enhancements for Content Strategy Autofill focusing on: learning from user acceptances, industry presets, constraint-aware generation, explainability, and RAG-lite context. It also captures the trade-offs for sectioned generation vs single-call generation.
Goals
- Increase accuracy, personalization, and trust without increasing UI complexity.
- Keep costs predictable while reducing timeouts and retries.
- Preserve user control: never overwrite locked/accepted fields without consent.
Single-call vs Sectioned Generation
- Single-call (current):
- Pros: 1 AI request, simpler orchestration.
- Cons: Larger prompt, higher timeout risk, brittle for structured JSON, hard to pinpoint failures.
- Sectioned (per category):
- Pros: Shorter prompts, better accuracy, quicker partial results, granular retries; lower latency per section; easier streaming (“Category X complete”).
- Cons: More calls; must cap/parallelize and cache to control cost.
- Recommendation: Hybrid
- Default: single-call for fast baseline; fallback/option: sectioned generation for users with large sites or when single-call fails/times out.
- Implement a server flag
mode=hybrid|single|sectionedand a per-user policy (feature flag).
Learning from Acceptances
- Data we already persist:
content_strategy_autofill_insights(accepted fields + sources/meta). - Learning policy:
- Build a per-user profile vector of “accepted values” and “field tendencies” (e.g., formats: video, cadence: weekly; brand voice: authoritative).
- During refresh:
- Use these as soft priors in prompt (“Bias toward previously accepted values unless contradictory to new constraints”).
- Prefer stable fields to remain unchanged unless explicitly unconstrained.
- Storage additions:
- Add fields to
content_strategy_autofill_insightsmeta:industry,company_size,accepted_at. - Maintain a compact, cached user profile (derived) for prompt injection.
- Add fields to
- Safety:
- Respect locked fields (frontend lock) → never modified by refresh.
Industry Presets
- Purpose: Cold-start quality boost.
- Source: curated presets per industry, company size, and region.
- Shape:
- Minimal key set aligned to core inputs (e.g.,
preferred_formats,content_frequency,brand_voice,editorial_guidelinestemplate).
- Minimal key set aligned to core inputs (e.g.,
- Retrieval:
- Endpoint: GET
/autofill/presets?industry=...&size=...®ion=...(cached).
- Endpoint: GET
- Merge policy:
- Apply only to empty fields; AI may override if constraints request.
Constraint-Aware Generation
- User constraints: budget ceiling, cadence/frequency, format allowlist, timeline bounds.
- UI:
- “Constraints” panel (chip-set) accessible from header/Progress area.
- Backend:
- Accept constraints in refresh request (query/body).
- Inject constraints into prompt header and soft-validate outputs.
- Validation:
- Enforce with server-side validators; warn if AI violates, and auto-correct when safe.
Explain This Suggestion (Mini-modal)
- Trigger: info icon next to each field.
- Content:
- Short justification text (one or two sentences), sources (onboarding/RAG docs), confidence.
- No raw chain-of-thought; ask model for a concise rationale summary that’s safe to expose.
- Backend payload additions:
- For each field:
meta[field] = { rationale: string, sources: string[] }(optional).
- For each field:
- Caution: redact sensitive content; keep rationale brief and non-speculative.
RAG-lite: Retrievable Context for Refresh
- Context sources:
- Latest website crawl snippets (top pages, headings, meta), recent analytics top pages (if connected), competitor headlines if available.
- Ingestion:
- Lightweight index (in-memory/SQLite) with page URL, title, summary; refresh on demand with TTL.
- Prompt strategy:
- Provide 3–5 top relevant snippets per category; keep token budget small.
- Controls:
- User toggle “Use live site signals” in refresh.
API Additions
- Refresh
- GET
/autofill/refresh/stream?ai_only=true&constraints=...&mode=hybrid&use_rag=true - Non-stream POST variant mirrors params.
- GET
- Presets
- GET
/autofill/presets?industry=...&size=...®ion=...→ returns compact preset payload.
- GET
- Acceptances (existing)
- POST
/{strategy_id}/autofill/accept→ persist accepted fields with transparency/meta.
- POST
UI Enhancements
- Per-field lock and regenerate
- Lock prevents overwrite; Regenerate calls sectioned refresh for that field’s category.
- Diff view on refresh
- Show before → after per field with accept/revert quick actions.
- Constraints chips
- Visible summary in header; edit inline.
- “Explain” modal
- Shows rationale and sources for the current value.
Observability & Metrics
- Track per-field fill-rate, violation corrections, latency (per section), AI cost per refresh.
- Alert on sudden drops in non-null field count or spike in violations/timeouts.
Rollout Plan
- Phase 1 (Low risk): presets + constraints + per-field lock, no sectioning.
- Phase 2: sectioned generation behind a feature flag; per-field regenerate.
- Phase 3: RAG-lite snippets and explain modal; start learning from acceptances in prompts.
- Phase 4: tune/fine-grain priors and add advanced validation rules per industry.
References
- Gemini structured output: https://ai.google.dev/gemini-api/docs/structured-output