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>
On the FX-8320 the first token took over a minute, and Sanic's default
60-second RESPONSE_TIMEOUT killed the stream before anything was sent.
Raise the timeout above the chat loop's own 600s hard ceiling, and emit
a status event before every model call so the stream is never silent —
for proxies and for the user staring at a slow machine.
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>
publish_ai_models.py produces the CDN artifacts for every AI_MODELS
entry — GPG-signed .meta4 sidecars (reusing the map trust chain) and
the signed ai/models.json manifest — and uploads them with s3cmd.
The bundled catalog now carries the real sha256/size of the two Qwen
GGUFs, which match Hugging Face's LFS records for the Apache 2.0
unsloth quants, and whose signed meta4s verify against the committed
public key.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Every MCP tool is now one /api/ai call with minimal text rendering;
the formatting, link building, and truncation logic it used to carry
untested lives (tested) in modules/ai. Tool names are unchanged for
external clients; a source-scan test guards the names and that the
proxy only reaches /api/ai. New lean blueprint endpoints back the two
tools that had no equivalent: GET archives/{id} (detail + snapshot
history) and GET videos/{id}/comments (paged author lines). Paged
content is read up to the old 50k-char bound. get_inventory_items now
takes the inventory slug (inventories are config-only; the old
numeric-id endpoint no longer exists). This also brings wrolpi_mcp
into version control (venv/ is already gitignored).
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>
install_llama_cpp.sh builds llama-server from the SHA-pinned v0.3.0
source tarball (single source of truth: on-device upgrade, Pi image,
Live ISO, and the new docker/llm container all call it; version-marker
makes re-runs no-ops). start_llama_server.sh refuses to start unless
ai.yaml enables AI and the active model file exists — which also covers
unmounted media. wrolpi-llm.service is installed but deliberately
never enabled at boot: the Controller starts it on demand (registered
in managed_services; the llm compose service follows the samba
pattern), and the API's perpetual idle-unload worker asks the
Controller to stop it after the configured idle minutes. llama-server
is localhost-only natively, internal-network-only in docker, and never
proxied by Caddy.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
ai.yaml is the source of truth (opt-in: enabled=False by default) with
active model, idle-unload minutes, context override, and tool-call
bound. The model catalog follows the map pattern: a bundled constant
keeps the Manage tab working offline, the CDN manifest (ai/models.json
+ detached GPG signature) wins when reachable. GGUF downloads clone
the map-extract flow — the per-file meta4 is GPG-verified once and the
same bytes handed to aria2c, so the verified hash is the enforced one;
unverified downloads are refused. Models land in <media>/ai/models/
(gated by the existing media_mounted download rules). Manage endpoints
(catalog with RAM-tier recommendation, settings) live under
/api/ai/manage/ which the future agent loop excludes from its tool
catalog.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Search the wrolpi-help markdown in-process: files are parsed on demand
and cached by mtime; title matches outrank headings outrank body. No
database table — the corpus is dozens of small files. Pages are read
as paged markdown. The docker api container gets a read-only bind
mount of the help docs at the native /opt/wrolpi-help location; a
missing help directory degrades to empty results with a message.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Proxy the Controller under /api/ai so the model sees one API: an
aggregate /status to start every diagnosis, lean service list, recent
service logs (lines clamped), and disks + SMART. One controller_client
helper resolves the Controller base URL (docker service name vs
localhost). Every proxy degrades to an `errors` entry when the
Controller is unreachable — never a 500.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Lean, flat, LLM-shaped read-only endpoints over the library: global,
video, archive, and doc search; paged captions, archive text, and zim
entries; collections, inventories, and a media-directory text reader.
Links are server-computed; big content is paged, never truncated.
This ports the formatting layer from wrolpi_mcp (untouched for now)
into tested code. First phase of ai.md.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
