mirror of
https://github.com/truespar/sentio.git
synced 2026-11-17 08:16:11 +00:00
README: drop the editorialising from the Jev section
This commit is contained in:
@@ -950,12 +950,6 @@ How much of a message leaves your server is your choice:
|
||||
| `"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).
|
||||
|
||||
|
||||
+1
-8
@@ -230,14 +230,7 @@ auto_respond = false
|
||||
base_url = "http://localhost:11434"
|
||||
model = "llama3"
|
||||
|
||||
# Only consulted when provider = "jev".
|
||||
#
|
||||
# Jev is closed-weight and hosted only - no self-host path, unlike Ollama. It
|
||||
# also proved unreliable in testing (429, 500, 503, 529; the last two are not
|
||||
# in their error reference), hence max_attempts below. A failed classification
|
||||
# is skipped, never fatal to delivery.
|
||||
#
|
||||
# To use it:
|
||||
# Only consulted when provider = "jev". To use it:
|
||||
#
|
||||
# [llm]
|
||||
# enabled = true
|
||||
|
||||
Reference in New Issue
Block a user