SDECB

Phase 2 Master Plan (Jan–Mar 2026) — Single Source of Truth

Purpose: This is the canonical “what’s true right now + what we do next” execution doc for Phase 2.
Non-goals: This file does not draft the actual AI policy or training curriculum content. It only defines process, structure, and delivery milestones.

---

1) Current reality (repo state)

This repo is the single source of truth for the SDECB project:

• Portal v2 lives in portal/v2/ and provides a clean UI for browsing/searching project content.
• Knowledge base (KB) is file-based and lives in knowledge_base/:
  • Canonical sources (raw exports / original materials)
  • Derived indexes (JSON) used by Portal v2 and future assistant retrieval

Canonical vs derived (rules)

Canonical (authoritative sources):

• knowledge_base/_sources/website/
  Raw website export (bilingual FR/EN, PDF-extracted markdown, entities/graph/catalogs as available).
• podcasts/episodes/ and podcasts/content/
  Episode artifacts (transcripts, metadata, content packs).
• interviews/
  Interview transcripts, summaries, and analysis.
• docs/
  Project documentation (implementation, analysis, governance process, training approach).

Derived (generated artifacts; safe to rebuild):

• knowledge_base/*/index.json and knowledge_base/unified_index.json
  Search/browse indexes generated by scripts.
• portal/v2/data/*.json
  Copies of KB indexes for static Portal v2 browsing.
• knowledge_base/_derived/embeddings/*.json
  Large embedding payloads (used for future semantic search; not required for Portal v2 MVP).

Rule of thumb: If it’s an index, it’s derived. If it’s the original text/audio export, it’s canonical.

---

2) Phase 2 goals (what “done” looks like)

Phase 2 is internal-first and bilingual (French/English):

• A usable KB that staff can search/browse via Portal v2, with clear source attribution.
• An internal assistant pilot that answers using retrieval + citations from the KB, with guardrails for sensitive content.
• Governance + training prepared as co-creation processes (structure and facilitation), not pre-written policy/curriculum.

Phase 3 stop/go gate: a public-facing chatbot only proceeds if internal assistant metrics and staff confidence meet agreed thresholds.

---

3) Operational workflow (how we keep it current)

A) Add / refresh content

• Website: drop/refresh export under knowledge_base/_sources/website/
• Podcasts: add episode artifacts under podcasts/episodes/ (FR first; EN when available)
• Interviews: add transcripts/summaries under interviews/

B) Rebuild indexes (derived)

Run the ingestion + index scripts in scripts/ to regenerate:

• knowledge_base/website/index.json
• knowledge_base/podcasts/index.json
• knowledge_base/interviews/index.json
• knowledge_base/unified_index.json

Then update Portal v2’s static copies under portal/v2/data/.

Reference docs:

• docs/technical/kb_architecture.md
• docs/technical/content_update_workflow.md
• docs/technical/kb_search_implementation.md

---

4) Internal assistant pilot (MVP architecture)

Principle

Answer with citations, always. The assistant is a navigation + synthesis layer, not a magical source of truth.

Retrieval scope (Phase 2)

• Default: include public and internal sources (website + podcasts + interviews)
• Always label sensitivity and include citations
• Refuse/redirect for requests that look like personal-data extraction or HR/performance/legal/medical advice

Local deployment model (minimal)

• A small local server provides:
  • static serving for Portal v2
  • POST /api/chat for the assistant UI
• Retrieval reads knowledge_base/unified_index.json and expands context from underlying markdown files when present.

Reference docs:

• docs/technical/portal_v2_architecture.md
• docs/technical/portal_v2_deployment.md

---

5) Governance + training (co-creation only)

We keep structure + process in-repo so the committee can co-create content with confidence:

• Governance hub: docs/governance/
  • policy_co_creation_process.md
  • policy_framework.md (skeleton only)
• Training hub: docs/training/
  • training_approach.md
  • curriculum_outline.md (structure only)

Portal v2 surfaces these as internal hubs for easy navigation.

---

6) Next milestones (execution order)

1. Documentation alignment: ensure repo status and docs match current reality (Phase 2 active, website+podcast assets present, indexes current).
2. Assistant backend MVP: local /api/chat server + env config.
3. Retrieval + citations: consistent scoring with Portal v2 search and stable citation objects.
4. Portal v2 Assistant page: bilingual UI, citations panel, sensitivity badges.
5. Pilot + eval: small bilingual eval set + success metrics + feedback loop.