API Reference

Assertions

Assertions define what properties each output must satisfy. Every assertion maps to one of four evaluation channels depending on the type.

Deterministic assertions

Evaluated in the SDK — no API call, no cost, <1ms.

valid-jsonFree

Output is valid JSON.

{ type: "valid-json" }

max-wordsFree

Output word count does not exceed the specified value.

{ type: "max-words", value: 400 }

min-wordsFree

Output word count is at least the specified value.

{ type: "min-words", value: 50 }

containsFree

Output contains the specified string (exact match, case-sensitive by default).

{ type: "contains", value: "Order #12345" }

not-containsFree

Output does not contain the specified string.

{ type: "not-contains", value: "competitor_brand" }

starts-withFree

Output begins with the specified string.

{ type: "starts-with", value: "Dear" }

languageFree

Output is written in the specified language (ISO 639-1 code).

{ type: "language", value: "en" }

matches-regexFree

Output matches the specified regular expression.

{ type: "matches-regex", pattern: "^\\d{4}-\\d{2}-\\d{2}$" }

Encoder assertions

Routed to DeBERTa-v3 or ModernBERT. ~$0.0001/call, 10–50ms.

not-toxicEncoder

Output does not contain hateful, threatening, obscene, or harassing content. Returns a toxicity score; passes if score < 0.1 by default.

{ type: "not-toxic" }
{ type: "not-toxic", threshold: 0.05 }  // stricter

no-piiEncoder

Output does not contain personally identifiable information (names, emails, phone numbers, addresses, SSNs).

{ type: "no-pii" }

sentimentEncoder

Output sentiment matches the specified value: positive, negative, or neutral.

{ type: "sentiment", value: "positive" }

Semantic assertions

Routed to the Glider judge model. ~$0.001/call, 100–500ms.

semanticSemantic Judge

The output satisfies the criterion described in natural language. The threshold field (0–1) sets the minimum score to pass. Defaults to 0.7.

1// These all route to the semantic judge:
2{ type: "semantic", criterion: "response demonstrates genuine empathy toward the customer", threshold: 0.75 }
3{ type: "semantic", criterion: "explanation is clear to a non-technical reader", threshold: 0.80 }
4{ type: "semantic", criterion: "code review is constructive and specific", threshold: 0.70 }
5{ type: "semantic", criterion: "answer is factually correct given the provided context", threshold: 0.90 }
6{ type: "semantic", criterion: "response maintains a professional but friendly tone", threshold: 0.75 }

Writing effective semantic criteria

→ Be specific. "Demonstrates genuine empathy" is better than "good response".
→ Describe the positive case, not the negative. "Stays on topic" not "doesn't go off topic".
→ Avoid compound criteria in one assertion — split them into two.

Common assertion patterns

patterns.ts

1// Customer support bot
2assertions: [
3  { type: "not-toxic" },
4  { type: "no-pii" },
5  { type: "max-words", value: 250 },
6  { type: "semantic", criterion: "acknowledges the customer's concern before offering a solution", threshold: 0.75 },
7  { type: "semantic", criterion: "offers a concrete next step or resolution", threshold: 0.80 },
8]
9 
10// Code review assistant
11assertions: [
12  { type: "valid-json" },  // if outputting structured review
13  { type: "min-words", value: 30 },
14  { type: "semantic", criterion: "identifies at least one specific improvement", threshold: 0.80 },
15  { type: "semantic", criterion: "tone is constructive rather than critical", threshold: 0.75 },
16]
17 
18// RAG answer quality
19assertions: [
20  { type: "max-words", value: 150 },
21  { type: "semantic", criterion: "answer is grounded in the provided context and does not hallucinate", threshold: 0.85 },
22  { type: "semantic", criterion: "answer directly addresses the question asked", threshold: 0.80 },
23]