Architecture · model and retrieval
French legal LLM: the model is only one layer of the product
A French legal LLM is a language model adapted or used for legal tasks in French; by itself it is neither an up-to-date database nor a reliable application. A complete solution may combine a model, retrieval from French sources, rules, permissions, interface, evaluation and human control. When comparing systems, ask which layer provides each function and how an answer is connected to its source.
Scope: French law · information, not legal advice
Model–RAG–application architecture cutaway
Six layers to inspect separately so model, corpus, interface and control are not confused.
- 01Language model
Generates and transforms text from parameters and context; ask about version, limits and usage terms.
- 02Legal corpus
Determines which French sources are available, for which dates and under which licences.
- 03Retrieval augmentation
Selects passages before generation; test recall, precision, filters and citation attachment.
- 04Rules and tools
Provide calculations, formatting, permissions or document access that the model should not improvise.
- 05Application
Organises conversation, documents, matters, exports and user roles.
- 06Evaluation and review
Measures failures on real uses and adds human validation where needed.
Choose a starting point without entering personal facts
Select a category only. No identity, date, amount, document or free-text question is included in analytics events.
Language model
Generates and transforms text from parameters and context; ask about version, limits and usage terms.
Keep the output, verify it against the listed sources, then obtain human review for any material decision.
What is a French legal LLM?
It is a language model used or adapted for French-language legal tasks; that does not guarantee an up-to-date French corpus or accurate citations. Check the source, its date and the full document before using this point in a decision.
What is the difference between an LLM and RAG?
The LLM produces text; RAG retrieves passages from a corpus and supplies them as context. Check the source, its date and the full document before using this point in a decision.
Must a model be trained on all of Légifrance?
Not necessarily. Well-designed retrieval can query authorised data without claiming the model memorises all French law. Check the source, its date and the full document before using this point in a decision.
Is an open-source model inherently more legal?
No. Model licensing alone says nothing about French corpus coverage, evaluation or source quality. Check the source, its date and the full document before using this point in a decision.
How is legal RAG evaluated?
Measure retrieval of the right authority, the useful passage and answer faithfulness to that passage separately. Check the source, its date and the full document before using this point in a decision.
Why does corpus date matter?
Legislation and official interpretation change; an answer should connect the rule to an applicable version and date. Check the source, its date and the full document before using this point in a decision.
Does fine-tuning replace retrieval?
No. It may change behaviour or style but does not automatically provide a current, verifiable source. Check the source, its date and the full document before using this point in a decision.
Where does Julie fit in this architecture?
Julie is a legal-assistance product; evaluate its published functions and method without attributing an undocumented model or corpus. Check the source, its date and the full document before using this point in a decision.
I am comparing AI architectures for French law. Help me separate model, corpus, RAG, tools, application, evaluation and human validation.