Topic Constraints¶

Prisma AIRS enforces hard limits on custom topic definitions. Daystrom handles these automatically so you don't have to worry about truncation or rejection.

Limits¶

Field	Max Length
Name	100 characters
Description	250 characters
Each example	250 characters
Examples count	2--5
Combined (description + all examples)	1000 characters

UTF-8 Byte Length

All length checks use UTF-8 byte length, not character count. Multi-byte characters (emoji, non-ASCII) consume more than one unit toward the limit.

Automatic Clamping¶

The LLM frequently exceeds the 250-character description limit — natural language tends to be verbose. The clampTopic() function enforces limits automatically after every LLM call:

graph TD
    A[LLM Output] --> B{Name > 100?}
    B -->|Yes| C[Truncate name]
    B -->|No| D{Examples > 250 each?}
    C --> D
    D -->|Yes| E[Truncate each example]
    D -->|No| F{Combined > 1000?}
    E --> F
    F -->|Yes| G[Drop trailing examples]
    G --> H{Still > 1000?}
    H -->|Yes| I[Truncate description]
    H -->|No| J[Done]
    F -->|No| J
    I --> J

The strategy is ordered by impact: drop trailing examples first, truncate the description only as a last resort. This preserves as much semantic content as possible.

Why post-LLM, not Zod

Clamping happens after the LLM generates output, not via Zod schema validation. The LLM needs freedom to produce natural descriptions, and the priority-based truncation logic is more nuanced than schema constraints can express.

Validation vs. Clamping¶

src/core/constraints.ts exports two approaches:

Function	What it does
Validation	Checks all limits, returns error strings. Non-destructive — useful for diagnostics and tests.
Clamping	Silently enforces all limits by truncating and dropping as needed. Used in the production pipeline.

Topic Name Locking¶

Info

After iteration 1, the topic name is locked — only the description and examples change in subsequent iterations. This prevents AIRS sync issues from name changes mid-run.