Knowledge bases and RAG
Page reviewed on
Retrieval-augmented generation: answers grounded in your documents, citing their sources, with an evaluation suite and access rights.
Retrieval-augmented generation (RAG) makes a language model answer from selected documents rather than from its memory alone. We design these knowledge bases so that each answer cites the passages it rests on.
What we do
- Selection and preparation of documents, chunking and indexing
- Answers that cite their sources, passage by passage
- Evaluation question set, with the expected answer for each, replayed at every version
- Access rights: users only get answers drawn from documents they are allowed to read
- Logging of questions, retrieved passages and answers
Formats
RAG projects fall under the Agentic systems and RAG offer:
- Scoping and architecture — 5 days.
- Prototype in real conditions — 15 to 25 days.
- Production roll-out, on quotation.
Deliverables
Documented indexing pipeline, evaluation suite, logs, access rules.
Who does the work
Fabrice Trollet, managing director of AGGIL, EFREI engineer. Background, publications and sources: References and sources.
Tax credit
AGGIL is approved for the French research tax credit (CIR) and innovation tax credit (CII) (2024-2026). For research or innovation operations entrusted to it only, expenses incurred with an approved body may be included in the client company's tax credit base, under the conditions and limits set out in Article 244 quater B of the French General Tax Code. The research and innovation tax credit offer