Ask a question

las ask ~/code/shop "Where is authentication done?"
las ask ~/code/shop "Onde usamos OpenAI?" --brief      # one line per evidence
las ask ~/code/shop "..." --no-llm                     # evidence only

Evidence comes first, the model second:

1. Deterministic search in claims, file names, function and class names, imports and routes — using the words of your question (accents and stop-words ignored) plus simple synonyms: authentication also looks for login, jwt, token, password, session…; payment for stripe, checkout, billing…
2. Excerpts: a few lines around each hit, best first. Product code ranks above tests and examples.
3. The model answers only from the excerpts, in the language of the question, citing file:line.
4. The validator removes every sentence that doesn't cite a line inside the excerpts.

≈ Answer drafted by AI from the evidence (model qwen2.5-coder:7b)
  ≈ Authentication is done in the login routes. [backend/app/api/routes/login.py:23]
  ≈ The verification of passwords is handled in the security module. [backend/app/core/security.py:30]

If nothing is found, the answer is "No evidence found about this" — with terms that do exist in the project as suggestions — and the model is never called. Asking about Stripe in a project without Stripe cannot produce an invented answer.

Without Ollama you get the list of evidence.