Future Roadmap

Future Roadmap — From Extractive Retrieval to Generative RAG

The following items are proposals, not currently implemented. The Mermaid designs in System-Architecture and Low-Level-Design represent the present source only.

Phase 1: Complete a real RAG pipeline

1. Document ingestion from files/URLs with parsers, chunking, metadata and tenant ownership checks.
2. Versioned embedding/index updates, deletion, reindexing and permission-aware retrieval.
3. LLM generation from retrieved context with grounded citations, token budgets and source provenance.
4. Answer-quality and retrieval evaluation, including groundedness, context precision/recall, and failure-case datasets.

Phase 2: Reliability and safety

• Identity provider integration; tenant and document authorization enforced beyond a single payload filter.
• Eliminate raw query echo or consistently sanitize response fields.
• Better PII detection, red-team evaluations, prompt-injection handling, documented data retention and audit trails.
• Shared semantic cache with invalidation/versioning and a multi-instance-friendly rate limiter.
• Distributed traces, structured metrics, health/readiness checks and alerts.

Phase 3: Scaling and advanced RAG

• Qdrant server/cluster hosting and backup policies.
• Batch inference or model serving, queueing and capacity tests.
• Hybrid dense + keyword retrieval, reranking and metadata filters.
• Streaming responses; model routing; access-controlled observability.
• Deployment manifests and CI/CD with security and integration gates.

Concrete next coding task

Add a new GenerativeRAGService module rather than changing _extractive_stub without tests. Define a retriever interface, prompt/context builder, LLM adapter, citation renderer and evaluation suite. The existing tenant-filtered retrieval would become one dependency of a fully generative pipeline.