docs(ai): AI Co-pilot system design — agentic RAG adapted for CA

[DeepakChander]

Summary

Adds docs/architecture/ai-copilot-system-design.md — the concrete blueprint for Phase 3.

Adapts the proven Agentic RAG pattern (Coordinator → Retrieval → Rerank → Context → Generation → Validation → Export) to CA-firm reality: regulatory citations, ICAI rules, UDIN/DSC gates, DPDP residency, and strict per-firm tenant isolation.

What's in it

• Architecture diagrams (online + offline pipelines)
• Tenant isolation model — per-firm Qdrant collection (structural, not metadata-filter)
• Two-collection retrieval — firm corpus + shared regulatory corpus, in parallel
• Full agent inventory including the new Compliance Critic agent (CA-specific)
• Citation model — every claim must resolve to a primary source (voucher ID / clause / SRN)
• Stakes tiering (Low / Medium / High / Blocked) replacing EdTech's "difficulty"
• UDIN / DSC gates — AI never marks output "final"
• Six CA use cases mapped to the same chassis
• Tech stack pinning: Qwen 2.5 (7B + 14B) via vLLM on Indian GPU host, BGE-M3, BGE Reranker, Qdrant, FastAPI
• Local dev with mocks — zero GPU
• Cost model + India residency table
• 12 failure modes + handling
• Phase 3A–3F sub-phasing plan with exit criteria

Test plan

• Markdown renders cleanly
• All internal links resolve
• Cross-references to existing docs (ai-copilot.md, tech-stack.md, security-and-data-residency.md, dpdp-act.md, icai-regulatory.md, pricing.md) are correct
• Linked from docs/README.md under Architecture
• CA advisor review (CODEOWNERS — pending merge)

Rollout notes

Docs-only change, no code or infra impact. Becomes the source-of-truth blueprint when Phase 3 build kicks off.

🤖 Generated with Claude Code