How the kit uses N-ATLaS

Every inference path in this repository is aimed at NCAIR1/N-ATLaS or one of the four NCAIR1 ASR repos. There is no fallback vendor.

flowchart LR
  subgraph apps [Apps]
    JS["n-atlas"]
    PY["natlas"]
    EX["examples"]
  end
  JS --> GW["Gateway serve"]
  PY --> GW
  EX --> JS
  EX --> PY
  GW -->|"POST /v1/chat/completions"| LLM["vLLM NCAIR1/N-ATLaS"]
  GW -->|"POST /v1/audio/transcriptions"| ASR["Whisper Small checkpoints"]

The diagram is Mermaid source. VitePress shows it as a code block; it is the same flow as the list below.

Chat

1. The SDK builds an OpenAI chat body. model defaults to NCAIR1/N-ATLaS and is rejected locally if it does not start with NCAIR1/.
2. Optional language is normalised to ha, ig, yo, or en and sent as a non-standard JSON field.
3. The gateway (prepare_chat_body in gateway.py) deletes language so vLLM does not 422 on an unknown key, and records the code for the log.
4. The gateway sets chat_template_kwargs.date_string to the UTC date in %d %b %Y form, unless the client already set one or NATLAS_INJECT_DATE_STRING=false.
5. vLLM renders the template shipped in the model's tokenizer_config.json. The launcher does not pass --chat-template.
6. The response is returned unchanged. Logs keep latency and token counts. Message text is not a log field.

translate, summarize, and detectLanguage are step 1 with a fixed English system prompt, then the same gateway path. voiceChat is transcription, then chat. The transcript is the user message. An empty transcript stops the chain.

Speech

1. The SDK sends multipart file, language, and response_format to POST /v1/audio/transcriptions.

2. normalise_language maps aliases onto ha / ig / yo / en.

3. settings.asr_models selects the repo:

   Code	Repo
   ha	NCAIR1/Hausa-ASR
   ig	NCAIR1/Igbo-ASR
   yo	NCAIR1/Yoruba-ASR
   en	NCAIR1/NigerianAccentedEnglish

4. ffmpeg converts the upload to 16 kHz mono. Audio longer than Whisper's 30-second window is split near silence (natlas_serve.audio).

5. Each chunk is run through that language's checkpoint. The texts are joined with spaces.

6. The HTTP response is { "text" } or the verbose fields. The transcript is not logged.

On Modal, steps 2–6 run in the gateway container and vLLM is a subprocess on the GPU. In Docker Compose, the gateway proxies the raw multipart body to the ASR container, which runs the same router.

What is intentionally absent

• No OpenAI, Anthropic, or Google SDK in any manifest. CI (.github/workflows/ci.yml, job guardrails) fails the build if one is added.
• backend="official" and backend="local" raise. They do not silently retarget.
• The playground directory is a separate app. It is not imported by serve/ or the SDKs. When it exists, it is expected to call this same gateway with the key kept on the server.

Keys

Secret	Where it lives	Who sees it
HF_TOKEN	Modal secret natlas-hf, or .env on a self-hosted box	The process that downloads weights. Not the browser, not the SDK
NATLAS_API_KEYS	Modal secret natlas-api, or the gateway environment	The gateway. Clients send one of those values as NATLAS_API_KEY
Request logs	stdout JSON	A k_ plus 12 hex chars of SHA-256 of the bearer token, not the token

Nothing in the docs site build reads these variables.