LLM Assignments

Under Settings → Intelligence → Kaiya & AI → LLM Assignments, admins can decide which saved provider configuration handles which piece of Kaiya's work. LLM Configuration registers the models; this page routes traffic to them. The split matters because Kaiya's tasks have different demands (a quick classification and a multi-step deep dive don't need the same model) so you can point fast work at one configuration and heavy reasoning at another.

Every dropdown on this page lists the reference names you created in LLM Configuration. A slot showing Select Configuration has nothing assigned yet.

Standard assignments

By default the page shows six broad categories. For most environments these are all you need to set.

General Intelligence: Everyday language work such as summarizing, classifying, pulling out entities, and working out what a question is asking. Suited to fast, context-aware answers across a wide range of requests.

Code Intelligence: Technical work, including SQL generation, Python scripting, and code explanation. This is the model behind analytical query generation and behind writing or debugging code.

Reasoning Intelligence: Multi-step reasoning, deep analysis planning, and layered contextual work. Use a stronger model here when answers need to connect findings across several stages.

Kaiya Architect: Conversational data preparation, schema design, and modeling suggestions when someone is building a Business View.

Apps: Content, logic, and configuration generated for Kaiya Apps.

Unstructured Content: Extracts structured information out of unstructured sources such as PDFs, emails, and free-text files.

Advanced Assignments

Turning on the Advanced Assignments toggle displays the full list of individual Kaiya tasks, and adds a search box above it. Type into the search box to filter the list by task name.

Use advanced mode when you need finer control than the six categories allow: routing exports to a cheaper model, giving SQL generation a stronger one, or pinning a specific task to a provider for cost or compliance reasons. Anything you leave unassigned falls back to the broad category behavior.

The six standard categories appear at the top of the advanced list, followed by these:

Task	What it does
Unstructured Reasoning	Deeper analysis of unstructured documents that need more comprehension than straightforward extraction.
Unstructured Fast	Speed-optimized handling of unstructured documents.
SQL Editor	Faster SQL generation that reuses SQL already held in the Semantic Cache.
Presentation Creator	Builds the PowerPoint structure from an uploaded template — slide structure, titles, and content layouts.
Web Search	Answers grounded in web search results, for questions that need current information and source citations.
Anchor Decision	Works out the anchors for a question from short-term memory, meaning the current conversation.
Orchestrator	Identifies intent at a high level using the configured workflow cards, then hands the work to the right sub-agent.
Time Filter Extractor	Pulls date and time expressions out of a question such as ranges, relative dates, and periods.
Time Filter Fallback	Steps in when the Time Filter Extractor can't process the query. Falls back to the coding model if nothing is assigned.
Long-Term Memory Retrieval	Finds and ranks relevant long-term memory entries so answers stay personalized and context-aware.
SQL Rephrasing	Rewrites the question for clarity and alignment before SQL is generated.
SQL Generation	Turns the question into SQL using the selected Business View's schema and context.
Preamble Generation	Writes the preamble that explains the thinking behind the SQL analysis.
SQL Summary	Plain-language summary of a SQL query and its results, so end users can read the outcome without reading the SQL.
Chart Config Generation	Builds the chart configuration and visualization spec from the query results and what the user asked for.
Insights Planner	Plans multi-step deep dives and decides the sequence of insights to generate.
Workflow Planner	Takes user-defined workflow steps and builds an execution plan for the Deep Insights engine.
Insights Summary	Detailed narrative summaries of analytical insights, calling out key findings and trends.
Python Retry	Diagnoses Python execution errors and writes corrected code so a failed step retries automatically.
Insights Concise Summary	Short summaries for cards, previews, and at-a-glance views.
Business View Summary	Describes a Business View's schema, relationships, and intended purpose in plain language.
Auto Business View	Picks or suggests the most relevant Business View for a given question.
Business View Validation	Checks Business View configuration — joins, column mappings, and data type consistency.
PowerPoint Export	Generates PowerPoint and PDF output for Deep Insights, Agentic Workflows, and similar, producing slide structures, titles, and content layouts.
PDF Export	Generates the storyline narrative and formatted content during a PDF export.
Export Citation	Checks the numbers, charts, and claims in an export against the source data and produces citation metadata.

Assigning and saving

Click any dropdown and pick a saved configuration. The one currently in use is marked with a checkmark. Only configurations saved in LLM Configuration appear here. If the provider you want is missing, add it here first.

Click Save to apply your changes, or Cancel to discard them and return the page to its saved state.