Forward Deployed Engineer, requirement 1

Retrieval-augmented generation

Grounded, cited answers over the customer's own documents, with access rules applied before retrieval.

The requirementBuild retrieval-augmented generation over a customer's own documents, so answers are grounded in their sources and cite them.

How I deliver it at a customer

  1. Scope the sourcesSit with the data owner, list the repositories, and agree what may be indexed and who may see it. Access rules are decided before any document is embedded.
  2. Chunk and embed inside the boundaryDocuments are split into passages and embedded. Luca runs embeddings on a local GPU by default, so customer text does not leave the environment; a hosted embedding model is the fallback.
  3. Hybrid retrievalEach passage is scored by semantic similarity blended with keyword overlap, then Maximal Marginal Relevance picks a diverse set, at most two passages per document, so one repeated paragraph cannot crowd out the rest.
  4. Grounded generationThe model answers only from the retrieved passages, cites them by number, and says so when nothing clears the relevance threshold instead of guessing.
  5. Access controlLuca keeps a separate document store per persona and checks the caller's role before retrieval, so a user never receives a passage they could not open directly.
  6. Measure before shippingA fixed set of customer questions with expected sources is re-run on every chunking, model or prompt change; citation accuracy and correct refusals are the pass criteria.

Tools I use

ToolWhat it does hereStatus
Luca Express RAG engineDocument store, embedding and query API inside the Luca workspace.Available
Luca OnEmbeddable website chat answering from a per-customer knowledge base.Available
Claude with MCPConnects the assistant to customer systems through open Model Context Protocol connectors.Available
Cloudflare Workers AIRuns this public demo at the edge: bge-base embeddings and a Llama instruct model.Demo only

Available means shipped and in use. Early access means built but not yet generally available. Demo only means it powers the public demo, not customer deployments.

Ask a question about the engineering work Gus IT has published. The demo retrieves from the 14 use cases on gusit.de with the hybrid method above and shows every score it used.

Open the live demo

Other answers

LLM applications in productionOne gateway for every model call, guardrails on both sides, and a measured choice of model per task.

Gus IT LLC. Contact gus@gusit.de. Claude is a product of Anthropic PBC. Cloudflare and Workers AI are trademarks of Cloudflare, Inc. Llama is a trademark of Meta Platforms, Inc. Other names are trademarks of their respective owners. Gus IT LLC is an independent engineering firm and is not affiliated with or endorsed by these companies.