diff --git a/README.md b/README.md index 2cdf7cf..d5c78c7 100644 --- a/README.md +++ b/README.md @@ -913,30 +913,51 @@ cheap ones could not decide: `score_llm_review_max`, 4.0-6.0 by default), so clear ham and clear spam never reach a model. -Four backends serve that third tier. Anthropic, OpenAI and Ollama classify a -message into a category. **TypeSafe Jev** (`provider = "jev"`) is a -structured-decision model: it answers typed questions with calibrated -probabilities rather than prose, so it is the only backend that also adjusts -the spam score. A confident verdict moves a borderline message clear of the -band in either direction; an unconfident one is discarded rather than applied -weakly. Being decision-only, it cannot write replies, so pairing it with -`auto_respond` is refused at startup. +Four backends serve that third tier: Anthropic, OpenAI and Ollama label the +message, and TypeSafe Jev also moves the score. -`body_mode` decides how much of a message leaves your server - `"none"` -(envelope and subject), `"preview"` (a bounded slice, the default) or -`"full"`. See `[llm.jev]` in [`config/oss.toml`](config/oss.toml) for the -weights, confidence floor and retry budget. +### TypeSafe Jev -Know what you are taking on before enabling it. Jev is closed-weight and -hosted only: there is no published weight file, parameter count or self-host -path, so unlike the Ollama backend you cannot run it yourself. During -development the service returned 429, 500, 503 and 529 - two of those -undocumented - so the provider retries transient statuses with exponential -backoff, and a failed classification is logged and skipped rather than -allowed to affect delivery. `base_url` is configurable if you would rather -reach the model through a gateway, or point at a compatible endpoint of your -own; neither is tested here. All of which is why the whole tier is opt-in and -the default provider is not Jev. +Jev is a structured-decision model - typed questions in, calibrated +probabilities out, no prose. It is the only backend that adjusts the spam +score rather than just labelling the message, so a confident verdict moves a +borderline one clear of the review band in either direction. An unconfident +verdict is dropped rather than applied weakly. + +```toml +[llm] +enabled = true +provider = "jev" +model = "jev-latest" +base_url = "" # empty means api.typesafe.ai +api_key_env = "SENTIO_JEV_API_KEY" +temperature = 0.3 # unused by Jev, but the section requires it +max_input_tokens = 2000 +classify_inbound = true +classify_outbound = false +auto_respond = false # required: Jev emits no text, so pairing it + # with auto_respond is refused at startup +``` + +About 660 input tokens per message and output is free, so roughly **$0.03 per +thousand messages** at current pricing. + +How much of a message leaves your server is your choice: + +| `[llm.jev]` `body_mode` | What is sent | +|---|---| +| `"none"` | Envelope and subject only | +| `"preview"` | Subject plus the first `preview_tokens` of the body (default) | +| `"full"` | Subject plus up to `max_body_tokens` | + +Two things to weigh before switching it on. Jev is closed-weight and hosted +only, so unlike Ollama you cannot run it yourself. And it was unreliable in +testing - 429, 500, 503 and 529, two of which are undocumented - so transient +failures retry with exponential backoff, and a classification that still +fails is logged and skipped rather than affecting delivery. + +Scoring weights, the confidence floor and the retry budget live in +`[llm.jev]` in [`config/oss.toml`](config/oss.toml). **Standards.** Core SMTP (RFC 5321/5322 and the ESMTP extensions), transport security (STARTTLS, MTA-STS, DANE, TLS-RPT), authentication (SASL, DKIM, SPF,