# Model registry — pin every revision by SHA or explicit tag. Never `latest`.
# Read by src/backend/config.py:load_models_config().
#
# After changing `embeddings`, rebuild the index:
#   uv run python scripts/build_index.py

llm:
  default:                      # Core generation LLM — Tiny Aya (3B, multilingual)
    backend: tiny_aya
    repo: CohereLabs/tiny-aya-global   # transformers (safetensors) repo — NOT the
                                       # -GGUF variant (GGUF has no HF tokenizer and
                                       # is only for the llama_cpp backend).
    revision: b2d281bab0beaecb3418d0c0dc53eaba90574011
    # NOTE: CC-BY-NC + gated. Accept the license on the model page and set the
    # HF_TOKEN Space secret. Runs locally via transformers on the Space (ZeroGPU);
    # without a token, generation falls back / errors locally (retrieval still works).
  swap_in:                      # One-line swap via get_llm_client("swap_in")
    backend: minicpm
    repo: openbmb/MiniCPM3-4B   # text LLM swap-in (adjust to a newer MiniCPM if desired)
    revision:

  # Future benchmark backend (GGUF over a llama.cpp OpenAI-compatible server).
  # Tiny Aya ships GGUF: CohereLabs/tiny-aya-global-GGUF
  #   llama_cpp:
  #     backend: llama_cpp
  #     base_url: http://127.0.0.1:8080/v1
  #     repo: CohereLabs/tiny-aya-global-GGUF

embeddings:                     # Query/passage embeddings for FAISS RAG
  repo: nvidia/llama-nemotron-embed-1b-v2   # Nemotron multilingual embed (1B, ungated)
  revision: 113abe4acafa848e77ead9c0623205e511932348                    # TODO: pin a commit SHA before deploy
  loader: sentence-transformers # SentenceTransformer(..., trust_remote_code=True)
  query_prefix: "query: "       # required by this model
  passage_prefix: "passage: "   # required by this model
  normalize: true               # L2-normalize so IndexFlatIP == cosine
  device: auto                  # auto → cuda on the ZeroGPU Space, cpu locally

parsing:                        # document parsing (OCR / table extraction)
  repo:                         # TBD
  revision:                     # TBD