LLMStep

Properties

context?

optional context: Record<string, unknown>;

Additional data to be inserted into the prompt. Accessible via Nunjucks variables.

Example

prompt: "Hello {{ name }}, your score is {{ score }}"
context: { name: "John", score: 42 }

---

execute()

execute: (reply) => Promise<unknown>;

Function to execute with the LLM's response. Use setProposedReply to use the LLM's output as the proposed reply. Or use combination of getProposedReply and setProposedReply to substitute parts of the string.

Parameters

Parameter	Type
reply	string

Returns

Promise<unknown>

Example

// Use LLM output directly as reply
execute: (reply) => messages.setProposedReply(reply)

// Substitute tokens in LLM output
execute: (reply) => {
  const withLink = reply.replace('<PAYMENT_LINK>', 'https://payment.example.com/123')
  messages.setProposedReply(withLink)
}

---

ignoreDirectives?

optional ignoreDirectives: boolean;

Exclude directives from message history passed to the LLM for this step

---

maxAttempts?

optional maxAttempts: number;

When provided, throws an error if the step is invoked more times than maxAttempts. Number of attempts taken is reset when shouldExecute returns false. Useful to limit rewinds by reviewers. NOTE that it doesn't work on steps without shouldExecute method.

---

model?

optional model: BasicModel;

LLM to use. Defaults to gpt-4o

---

name

name: string;

Step name for debugging

---

onError()

onError: (error) => Promise<unknown>;

Error handler called if an error occurred during LLM API call or in execute function

Parameters

Parameter	Type
error	string

Returns

Promise<unknown>

---

prompt

prompt: string | File;

Prompt can be a simple string or a link to a file, loaded with loadFile function which takes a path to the file relative to src/use-cases directory. Should be Nunjucks-compatible.

---

runIf()?

optional runIf: (messages) => boolean | Promise<boolean>;

Determines if the step should be run or not

Parameters

Parameter	Type
messages	Conversation

Returns

boolean | Promise<boolean>

---

schema?

optional schema: ZodType<any, ZodTypeDef, any>;

Schema for structured LLM output using zod https://zod.dev/ library.

---

shouldExecute()?

optional shouldExecute: (reply) => boolean | Promise<boolean>;

Check a condition, whether the execute function should run or not

Parameters

Parameter	Type
reply	string

Returns

boolean | Promise<boolean>

Deprecated

use runIf to check if the step should be run, use if in execute to check if it should be executed