AI & Automation

AI Skill Master

An AI Skill defines one AI job: what it does, which provider and model it uses, the instructions it is given, and what happens to the result. Skills are the only place this is configured — no prompt, model, or endpoint is hard-coded in Quasar. The Skill Type decides which fields matter: a Vendor Invoice skill uses the Input and Instructions tabs, while a Voice Command skill uses the Voice & Chat tab. Access from System › Setup › AI Skill Master, or from the AI Skill List.

Skill Types

Type	Runs	Purpose
Vendor Invoice (Receive)	Server	Reads vendor invoice PDFs from a watch folder and posts or stages a Vendor Invoice.
Bank Reconcile	Server	Reserved for bank statement import. Not yet implemented.
Voice Command	Workstation	Spoken commands. Uses the provider's Voice Command URL to turn speech into text, which Quasar then acts on.
Voice Chat	Workstation + Server	Spoken conversation with the AI. Uses the same Voice Command URL speech-to-text stream as voice commands; each utterance is answered by the Quasar server and shown as text in the AI Assistant panel (no spoken replies).
Text Chat	Server	Typed conversation in the AI Assistant panel. The workstation sends the text to the Quasar server, which resolves the skill and provider and returns the answer.
Data Entry / Report Query / Email Parse / Custom	—	Reserved for future jobs.

When more than one skill of the same type exists, they are tried in Number order.

General Tab

Number

The skill number. Also sets the order skills of the same type are tried in.

Name

The descriptive name shown wherever the skill is selected.

Description

Free-form notes about what this skill is for.

Skill Type

What the skill does. See the table above.

AI Provider

The AI service this skill uses (lookup). The provider must supply the URL the skill type needs.

Model

The model to use. Must be one of the provider's Allowed Models. Leave empty for voice skills to use the first voice-capable model the provider offers.

Vendor

Default vendor for invoices imported by this skill. Vendor Invoice skills only.

Account

Default account for bank reconciliation skills.

Media Type

Media type used when the source PDF is attached to the posted transaction.

Posting Mode

What the job does with a successfully extracted invoice.

Complete and Post

Post the Vendor Invoice automatically when every line resolves to an item.

Save Suspended

Create the Vendor Invoice but leave it suspended for someone to check and post.

Stage for Review

Stage the invoice in AI PDF Invoices for line-by-line review. The safest setting, and the recommended default.

Update Costs

Update the item's cost record when the invoice cost differs from the stored cost.

Create Price Batch

Create a price batch for items whose cost changed, so selling prices can be reviewed against the new cost.

Input Tab

Where the job's documents come from. Used by server-side skills; ignored by voice and chat skills.

Input Source

Watch Directory, Manual Upload, Email, or API.

Watch Directory

Folder the job polls for new documents. Processed files are moved to a processed subfolder, and failures to an error subfolder.

File Pattern

Which files to pick up, e.g. *.pdf.

Poll Interval (ms)

How often the watch directory is checked.

Processing Order

Order this skill runs in relative to other server-side AI skills.

Sample Directory

A sample document used by Analyze Sample to work out the column layout.

Browse

Choose the sample file.

Analyze Sample

Sends the sample to the provider and reports the columns and fields it found, so mappings can be set up without processing a live invoice.

Analysis Status

Progress and result of the last analysis.

Analysis Result

The raw analysis returned by the provider.

Generate Instructions

Builds a starting extraction prompt from the analysis result, which can then be edited on the Instructions tab.

Column Mappings

Maps a column heading on the vendor's invoice to the Quasar field it feeds. Each row has the PDF column name and optional notes.

Instructions Tab

System Prompt

The instructions given to the model on every request. For voice and chat skills this sets the assistant's behaviour; for extraction skills it sets the ground rules.

Extraction Prompt

The task instructions and the JSON schema the model must return. Extraction skills only.

Validation Rules

Notes describing checks to apply to the extracted data before it is used.

Ignore Columns

Columns on the source document the model should skip.

Ignore Text

Text patterns the model should skip, such as promotional footers or page furniture.

Responses

Sample or recorded responses, used to check the prompt behaves as intended.

Voice & Chat Tab

Settings for the workstation assistant. Applies to Voice Command, Voice Chat, and Text Chat skills.

Hotkey

Key that starts and stops listening. F7 by default.

Language

Spoken language for transcription, or Auto-detect.

Auto Submit

Act on a command as soon as it is recognised, without waiting for confirmation.

Fallback to REST

If the streaming connection cannot be opened, fall back to sending a recorded clip over the REST API.

Audio Format

Microphone capture format sent to the provider, normally PCM 16-bit.

Audio Sample Rate

Capture sample rate in Hz. 16000 suits speech recognition.

Allowed Actions

Which actions the assistant may perform — focus, fill, fill_range, search, help, open_screen, clear. Every action is still checked against the user's security rights when it runs; this list narrows what the assistant may ask for in the first place.

Actions Preview

Read-only summary of the actions currently allowed.

Provider Warning

Appears when the selected provider lacks the URL this skill type needs, for example a Voice Command skill on a provider with no Voice Command URL.

Load Default Prompt

Fills the System Prompt with the shipped voice command template as a starting point.

Set Active

Makes this the skill the workstation assistant uses for its type. The choice is stored per workstation, so set it once on each workstation that uses voice.