Changelog

The 1B model was inventing intent IDs like 'call' instead of using

tasks.create_task. Added a final IMPORTANT reminder listing all valid

intent_id values and 'Do NOT invent new IDs' to anchor the model to

the available actions. Bumps desktop to v1.41.8.

Example 4 was teaching the model that 'call the dentist to book

appointment' is a calendar event, but it's actually a task (an action

to perform). Rewrote examples to explicitly distinguish: confirmed

meetings at specific times = calendar.create_event, things the user

needs to DO (call, send, buy, book) = tasks.create_task. Added

explicit guidance at the top of the prompt for this distinction.

Bumps desktop to v1.41.7.

The 1B model was listing irrelevant intents (create_note from a note,

create_contact from appointment text) without extracting slots. Key

changes: (1) explicit instruction "most text has 0 or 1 actions, be

selective", (2) require filled slots, (3) added Example 4 showing a

mixed sentence where only one part is actionable — teaches the model to

be selective. Removed entity title/type from user prompt to avoid

confusing the model with source metadata. Bumps desktop to v1.41.6.

The 1B local model was ignoring the available actions list and inventing

its own intent IDs from the text content. Added three few-shot examples

(calendar event, task, and no-action) so the model sees the expected

output format with real intent IDs. Reduced MaxTokens to 384 and

temperature to 0.2 for tighter output. Bumps desktop to v1.41.5.

The previous prompt was ~1400 tokens with verbose slot listings, JSON

schema examples, and multiple rules — too much context for a 1B model.

The model was ignoring the actual content and picking random intents.

Rewrote to a compact format: one line per intent with just the ID,

description, and required slot names. Single-line JSON example. Reduced

MaxTokens from 1024 to 384 and temperature from 0.3 to 0.2 to keep

output tight. Added debug logging for raw AI responses. Bumps desktop

to v1.41.4.