flemming-it 4d6acab6a1

CI / Linux x86_64 (Forgejo) (push) Failing after 2s

Details

sign-bundle / sign (push) Failing after 1m48s

Details

feat: api accepts 'vllm' as a first-class value

'openai' reads as the cloud company — an operator running
self-hosted vLLM should be able to write what they mean. The new
value speaks the identical OpenAI wire (a guard test asserts the
emitted request stays byte-identical to api: openai, so the alias
can never drift into a dialect) but yields vLLM-specific error
hints (vllm serve, port 8000) instead of generic OpenAI prose.
Manifest + docs name the value in DE and EN.

Signed-off-by: flemming-it <sf@flemming.it>

2026-08-20 17:13:42 +02:00

3.8 KiB

Raw Permalink Blame History

text.translate

LLM-backed translation between languages (Ollama, vLLM, any OpenAI-compatible server, or Anthropic). Sends source text to a configured LLM endpoint with a translation system prompt and emits the translated text plus audit-grade model-provenance fields.

Capability

text.translate@0.1.0

Inputs

Name	Type	Description
`text`	text	Source text to translate.
`target_language`	text	Target language in plain English (e.g. `German`, `French`, `ja-JP`).
`source_language`	text	Optional source-language hint. Empty = let the model auto-detect.
`endpoint`	text	Chat endpoint URL matching the selected `api` (Ollama `/api/chat`, OpenAI-compatible `/v1/chat/completions` — vLLM etc., Anthropic `/v1/messages`).
`api`	text	Optional wire format: `ollama` (default), `vllm`, `openai`, `anthropic` (`vllm` = the OpenAI wire with vLLM-specific hints).
`model`	text	Model identifier the endpoint serves (e.g. `qwen2.5:14b`).
`api_key`	text	Optional bearer token for cloud-hosted endpoints.

Outputs

Name	Type	Description
`translation`	text	The translated text.
`source_language`	text	Echo of the input value (or detected language when auto).
`target_language`	text	Echo of the input value.
`model_endpoint`	text	URL the translation was generated against.
`model_name`	text	Model identifier as supplied to the LLM API.
`model_digest`	text	SHA-256 digest of the served model — Ollama only, empty elsewhere.

The three model_* outputs are the audit primitive: every translation in the flow log can be matched to the exact model hash that produced it, even months later.

Permissions

permissions:
  - "net: localhost"
  - "net: 127.0.0.1"
  - "net: api.openai.com"
  - "net: api.anthropic.com"

The default policy allows local Ollama and the two major hosted providers. Operators with a private LLM endpoint add their hostname to the operator-policy ceiling; the module's permissions are enforced against that ceiling at install time.

Built for regulated environments

The Ollama endpoint defaults to a local daemon — source text never leaves the host. Cloud providers are opt-in via the endpoint URL + operator policy. The model_digest field is the cryptographic root-of-trust for translation provenance: two outputs with the same digest came from the same model weights, period.

Limits in v0.1.0

Single-shot translation. Long-document chunked translation with overlap-stitching is the v0.2.0 plan.
No glossary / terminology-list support. Operators currently rely on the model's general knowledge; a glossary input is on the v0.2.0 roadmap.

Example flow

name: translate-to-german
inputs:
  source: text
steps:
  - id: translate
    use: text.translate@^0
    with:
      text: $inputs.source
      target_language: "German"
      endpoint: "http://localhost:11434/api/chat"
      model: "qwen2.5:14b"
outputs:
  translation: $translate.translation
  audit_model: $translate.model_digest

Build

cargo build --release --target wasm32-wasip2
# Output: target/wasm32-wasip2/release/text_translate.wasm

3.8 KiB Raw Permalink Blame History