LiteLLM

Djelia works as a standard OpenAI-compatible provider. Prefix the model with openai/ and point api_base at the compatibility surface.

import litellm

litellm.completion(
    model="openai/banjugu-1",
    messages=[{"role": "user", "content": "Bonjour"}],
    extra_body={"djelia": {"source_language": "fra_Latn", "target_language": "bam_Latn"}},
    api_base="https://api.djelia.cloud/openai/v1",
    api_key="YOUR_DJELIA_API_KEY",
)

litellm.transcription(
    model="openai/sunjata-1",
    file=open("audio.mp3", "rb"),
    api_base="https://api.djelia.cloud/openai/v1",
    api_key="YOUR_DJELIA_API_KEY",
)

litellm.speech(
    model="openai/jifili-1",
    input="Aw ni ce",
    voice="moussa",
    api_base="https://api.djelia.cloud/openai/v1",
    api_key="YOUR_DJELIA_API_KEY",
)

Proxy

The same three entries go in config.yaml:

model_list:
  - model_name: djelia-translate
    litellm_params:
      model: openai/banjugu-1
      api_base: https://api.djelia.cloud/openai/v1
      api_key: os.environ/DJELIA_API_KEY
  - model_name: djelia-asr
    litellm_params:
      model: openai/sunjata-1
      api_base: https://api.djelia.cloud/openai/v1
      api_key: os.environ/DJELIA_API_KEY
  - model_name: djelia-tts
    litellm_params:
      model: openai/jifili-1
      api_base: https://api.djelia.cloud/openai/v1
      api_key: os.environ/DJELIA_API_KEY

Cost tracking

Djelia models are not in LiteLLM's built-in cost map, so response_cost comes back as None until you register them. Because usage reports characters and translation is metered per input character, a per-token price set to the per-character price yields the exact charge:

litellm.register_model({
    "openai/banjugu-1": {
        "input_cost_per_token": 0.00005,   # $/character
        "output_cost_per_token": 0.0,      # output is not metered
        "litellm_provider": "openai",
        "mode": "chat",
    }
})

Transcription is metered per second of audio, which has no equivalent in LiteLLM's token-based cost model. Use Djelia's own usage records for ASR spend.