Response-quality work, all aimed at giving the model enough context to
answer accurately:
- Every chat's system prompt now carries a live library map (today's
date, counts, and the exact names of channels, domains, zims, and
inventories, capped and cached 60s). Exact spellings are in context
before the first token, and the list_collections round-trip — minutes
on slow hardware — is usually unnecessary.
- The request is budgeted against the model's real context: oldest
turns are trimmed first, the current question is never dropped, and
an over-budget turn gets a clear "start a new conversation" error
instead of a silently truncated prompt. cache_prompt reuses the KV
cache for the stable prefix, which is most of the latency on a Pi.
- context_size plumbing fixed: selecting a model adopts its catalog
default (the 4B was running at 8k instead of 16k), and llama-server
runs --parallel 1 so the single user gets the whole context.
- temperature is configurable in ai.yaml (default 0.3 — small models
call tools far more reliably near 0.2-0.3 than at server defaults).
- One worked tool-use example per mode is prepended to every
conversation; small models imitate far better than they obey.
- scripts/ai_eval.py scores the assistant against a live WROLPi on
mechanical checks (finished, called the right tool first, included a
WROLPi link, no external URLs) so prompt changes stop being validated
by anecdote.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Live testing surfaced three ways small models went wrong summarizing a
library: they gave up after one empty search (misspelled channel names
found nothing), so zero-result searches now carry a retry hint and the
Research prompt requires list_collections before channel/site questions
and search_videos(channel_id) before claiming a channel is summarized;
they presented the familiar YouTube source URL instead of the WROLPi
link, so listings omit the source URL (archive detail keeps it); and
they invented links for channels, so channel collections now carry
their real /videos/channel page link. Search descriptions state that
search_str is optional for browsing the newest items.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three findings from running Qwen3-1.7B against the docker stack:
the llm image lost libgomp when the build tools were removed
(llama-server needs the OpenMP runtime); the Help prompt let the
small model answer from memory instead of the docs, so it now
requires search_help before any answer; and help search results
called the slug "doc" while get_help_doc's argument is "slug", so
the model invented slugs — the names now match.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Tool definitions are generated in-process from the /api/ai blueprint's
own OpenAPI operations (operationId = tool name, description = tool
description, dataclass bodies + declared parameters = one flat JSON
schema) — the spec IS the catalog, nothing to drift. Manage endpoints
never become tools. A mode is a server-enforced system prompt + tool
allowlist (Help / Research / System per ai.md); the executor can only
reach allowlisted /api/ai endpoints, path parameters are encoded and
traversal-rejected so a crafted argument cannot escape an endpoint.
POST /api/chat streams SSE (status/token/tool_call/tool_result/done/
error), starts llama-server on demand, validates tool args server-side
with instructive errors the model can retry on, caps tool results with
a paging hint, and bounds the loop by ai.yaml max_tool_calls plus a
hard timeout. Caddy passes the stream unbuffered (flush_interval -1)
in both deployments.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
