A marketing lead asks for the minimum safe RAG architecture for versioned documents
FitTechnology6335 · reddit · 2026-07-29
A marketing lead asks for the minimum safe RAG architecture for versioned internal docs
The author is building an internal knowledge system for a construction company and needs the LLM to work with about 60 documents spanning content history, analyses, instructions, examples, and multiple versions of some files. The current MVP is deliberately simple: a private portal gives users a universal prompt plus a temporary read-only link, and they paste both into ChatGPT, Claude, or Gemini.
The problem is that the corpus can exceed context limits and, more importantly, contains active, superseded, historical, and experimental versions. The author no longer wants the model to decide which document is authoritative.
The post compares three approaches:
- A. Curated active bundle + manifest: generate a clean package with only active documents plus a manifest describing versions and relationships, while archiving historical material separately.
- B. Metadata-filtered retrieval service: tag documents with documentid, version, status, supersedes, and effectivedate, then filter authoritative docs before semantic retrieval so the model never chooses the version.
- C. Custom chat/RAG application: turn the portal into the actual interface, with retrieval and context selection handled in the app rather than by prompting ChatGPT/Claude/Gemini directly.
The author asks whether A is a safe MVP or whether B is already the minimum once version authority matters, and invites architecture warnings and alternatives.
More from Companies & People
- User says Anthropic’s AI support bot missed a basic support request — RichmanRonald · 2026-07-29
- OpenAI’s Codex Ambassador program sends swag and community updates — paw_lean · 2026-07-29
- Chip Huyen’s AI Engineering GitHub repo bundles summaries, notes, and prompt examples — mdancho84 · 2026-07-29
- Google DeepMind opens staff scientist role for Gemini long-context and memory — LucaAmb · 2026-07-29
- Why enterprise AI projects lose money — and where the real budgets are — yangyi · 2026-07-29
- Stanford AO Summit spotlights autonomous organizations, from agent swarms to AI villages — allisondman · 2026-07-29