id: WO-0008
title: "Maya — Conversational KB Talent (RAG-as-Wiki)"
requestor: Oscar (CEO)
assignee: Pablo (Production Line Architect)
participants: [Riley (R&D Analyst), Camille (Customer Success), Kai (Knowledge Manager), Ivan (Infrastructure), Diego (Deployment)]
status: in-progress
jira_key:
created: 2026-05-12
started: 2026-05-26
completed:
decision_ref: TFD-0023
impacts:
process: client request intake & deflection (CON-0006)
capability: replaces Confluence wiki with conversational RAG; bundleable with every delivered digital talent
roles: [customer-success, knowledge-manager]
client: Internal first (factory + STM pilot), then bundled with each delivered talent
product_type: Conversational KB digital talent (RAG over markdown + HTML corpus)
production_line: digital-talent
language: French (primary) / English
references: [CON-0006, RD-0003, RD-0028]
Request
Build Maya (Memory & Answers) — a conversational knowledge base talent that replaces the JSM-Confluence deflection pattern with native RAG. Maya reads a corpus (markdown decisions, request outputs, HTML deliverables, lessons learned), answers client questions conversationally with citations, and routes unresolved cases to ticket creation.
Two distinct deployments:
Maya-Factory — internal: reads talent-factory/ corpus (decisions, requests, KB learnings). Used by Camille for client triage, by the team for institutional knowledge retrieval.
Maya-Client (per client) — bundled with each delivered digital talent. Reads that client's delivered solutions + KB. Becomes the deflection layer in the client portal.
Sibling pattern to dt-coach, negotiation-coach, fitness-coach — but the first factory infrastructure talent that is also a client deliverable.
Strategic rationale
Cannibalize Confluence: JSM-Confluence deflection (CON-0006 Stack A) costs money + is keyword-based + can't render rich HTML. Maya beats it on every axis: native FR/EN, semantic understanding, conversational, free corpus (git), bundleable.
Eat your own dog food: the factory builds digital talents that read docs and answer; Maya is that pattern turned inward.
Product differentiation: every delivered talent ships with its own Maya. Clients get a chatbot trained on their solutions. Hard for competitors to match.
Aligned with feedback_delivery-model-foundry-not-hosted: Maya is packaged as a digital talent, not a hosted service. Customer runs it independently with their LLM of choice.
Client Profile (M1 — internal pilot)
Field
Value
Client
Talent Factory (internal)
Segment
Camille (CS), Riley (R&D), the CEO
Need
Conversational retrieval over factory KB; basis for client-facing deployments
Language
French primary, English
AI platform
Claude API (Haiku for retrieval, Opus/Sonnet for synthesis)
Product Definition
#
Capability
Description
1
Corpus ingestion
Index a directory of markdown + HTML + PDF into a manifest (titles, summaries, paths). No vector DB required for <10k docs.
2
Conversational retrieval
Multi-turn dialogue; understands FR/EN mixed; clarifies ambiguous queries before answering
3
Citation by paragraph
Every claim cites the source file + section (Claude API citations native)
4
Deflection → ticket handoff
When Maya can't answer or user says "ouvre un ticket", capture the conversation context and route to JSM (or Telegram for Maya-Factory)
5
Embeddable widget
Single <script> widget that drops into Astro/Vercel pages, JSM portal, Teams app
6
Bilingual native
FR/EN mixed input handled without language switching; output matches user's last language
7
Per-deployment corpus
Each Maya instance is configured with corpus_path — factory uses talent-factory/, STM Maya uses OneDrive-STM/agent-ea/, etc.
8
Re-indexing on commit
Hook on git commit updates the manifest. Always fresh against the corpus.
Out (v1): Vector DB (overkill for current volume), authentication (handled by the host portal), analytics dashboard (handled by JSM/portal)
Compliance: Each Maya instance only reads its configured corpus. No cross-tenant leakage. Citations always include source path.
Acceptance Criteria
Given a question in French about the factory R&D pipeline, when asked through the widget, then Maya answers with citations to the right files in departments/executive/rd-analyst-riley/ and company/decisions/TFD-0012*
Given a question in English mixed with French terms (typical CEO style), when asked, then Maya parses and answers without forcing language switch
Given a question Maya cannot answer from corpus, when reached, then Maya proposes "create a ticket" and forwards conversation context to JSM (or Telegram for Maya-Factory)
Given a corpus update (git commit), when triggered, then the manifest re-indexes within 60s and new content is searchable
Given a Maya-Client deployment for STM, when STM user queries, then only STM corpus is read (no factory-internal leakage)
Given any answer, when delivered, then citations link to the actual source file (clickable in widget)
Definition of Done
All acceptance criteria pass
Maya-Factory deployed to factory team (CEO + Camille + Riley access)
One Maya-Client instance configured for STM corpus as POC
Deployment package: standalone digital talent with corpus_path env var, deployable to client repo
Widget embeddable in JSM portal (POC) and beta-portal (production)
QA certification completed (qa-certification.md)
Deflection handoff to JSM verified end-to-end
Documentation: client integration guide (how to wire Maya into their portal)
Success Criteria
30-day: Maya-Factory answers ≥70% of internal team queries correctly without fallback to manual search. Maya-Client (STM) handles ≥3 real STM questions during pilot.
90-day: Maya bundled with at least 2 delivered talents (STM agent-ea + 1 other). Confluence usage in factory drops or stops entirely. Measurable deflection rate on JSM portal (% tickets prevented).
Agent Context
References
CON-0006 — Client request intake lifecycle (Maya is the deflection layer in that architecture)
RD-0003 synthesis-strategique — §3 "What the video omits" mentions memory systems and bilingual as factory assets
project_framework-library — Maya may need to read framework docs as part of corpus
feedback_documentation-format — HTML deliverables in corpus must remain readable by Maya
feedback_delivery-model-foundry-not-hosted — Maya packaged for customer execution with LLM flexibility
Constraints
Follow production line stages: production-lines/digital-talent/
Naming: WO-PROD-{NNN}-{slug} per TFD-0009
LLM flexibility per foundry model: Maya must work with Claude API by default but accept Anthropic-compatible local model endpoints (per model-config-pattern). No hard-coded model.
Bilingual FR/EN support; French primary for STM
No vector DB in v1 — manifest-based retrieval + Claude long context. Re-evaluate at 10k docs.
Citation format must be machine-parseable so the widget can render clickable source links
Verification
Run: standard digital-talent QA gate (8/8 functional, ≥5/6 edge cases)
Additional: deflection routing test (Maya → JSM ticket creation with conversation context preserved)
Expected: QA pass, deployment manifest, internal pilot live, STM POC live