LLM - Weights & Biases Documentation

An LLM call. Emits a chat span with gen_ai.* attributes. Created by weave.startLLM() (or turn.startLLM()) and terminated with end(). Only one LLM may be active in an async context at a time; nest tool/subagent calls under it via startTool / startSubagent. Populate inputMessages / outputMessages / usage / reasoning directly, or via the helper functions (output, think, attachMedia, record). All recorded data is flushed to the span at end().

Examples

const llm = weave.startLLM({model: 'gpt-4o-mini', providerName: 'openai'});

try {
  llm.inputMessages = [{role: 'user', content: prompt}];
  const resp = await openai.chat.completions.create({...});
  llm.output(resp.choices[0].message.content ?? '');
  llm.record({usage: {inputTokens: resp.usage?.prompt_tokens}});
} finally {
  llm.end();
}

const llm = weave.startLLM({
  model: 'gpt-4o-mini',
  providerName: 'openai',
  systemInstructions: ['You are a helpful weather bot.'],
  startTime: new Date('2026-05-29T10:00:00.000Z'),
});

try {
  // ... call the LLM, populate llm.outputMessages / usage ...
} finally {
  llm.end();
}

Extends

SpanBase

Properties

inputMessages

inputMessages: Message[] = []

Input messages sent to the model. Flushed to gen_ai.input.messages on end().

model

readonly model: string

outputMessages

outputMessages: Message[] = []

Assistant messages returned by the model. Flushed to gen_ai.output.messages on end().

providerName

readonly providerName: string

reasoning

optional reasoning: Reasoning

Chain-of-thought content. Folded into the last assistant message as a ReasoningPart at serialization time.

usage

usage: Usage = {}

Token counts and cache stats. Flushed to gen_ai.usage.* on end().

Methods

addEvent()

Deprecated. Record this data via setAttributes instead. OpenTelemetry is phasing out the Span Event API (Span.addEvent). This method still works and existing span-event data stays valid. See https://opentelemetry.io/blog/2026/deprecating-span-events/

addEvent(name, attributes?, startTime?): this

Add a named event to the span. Useful for marking non-span moments such as context compaction, tool-loop detection, or guardrail trips. Warns and no-ops after end(). Mirrors OTel Span.addEvent.

Parameters

string

required

Attributes

TimeInput

Returns

this

Example

span.addEvent('context_compacted', {removedMessages: 12});

Inherited from

SpanBase.addEvent

attachMedia()

attachMedia(opts): this

Stage a media attachment for the LLM call. Pick exactly one of content (inline base64 bytes), uri (URI reference), or fileId (pre-uploaded file id). The attachment is glued onto the last user message in inputMessages on end().

Parameters

{…} | {…} | {…}

required

See Modality, Modality, Modality.Full type:

{ content: string; mimeType: string; modality: Modality; } | { modality: Modality; uri: string; } | { fileId: string; mimeType?: string; modality: Modality; }

Returns

this

attachMediaUrl()

attachMediaUrl(url, opts): this

Convenience for attachMedia({uri, modality}).

Parameters

string

required

object

required

Hide properties

Modality

required

See Modality.

Returns

this

end()

end(opts?): void

Flush accumulated state and close the span. Idempotent. Pass error to mark failed; pass endTime to backdate the close.

Parameters

object

Hide properties

TimeInput

Error

Returns

void

output()

output(content): this

Append an assistant message to the response.

Parameters

string

required

Returns

this

record()

record(opts): this

Bulk-set any subset of the mutable fields. Replaces (does not merge). Useful for assigning everything at once after a provider call returns.

Parameters

object

required

Hide properties

string[]

Message[]

See Message.

({…} | {…} | {…})[]

See Modality, Modality, Modality.Full type:

({ content: string; mimeType: string; modality: Modality; } | { modality: Modality; uri: string; } | { fileId: string; mimeType?: string; modality: Modality; })[]

Message[]

See Message.

string

Reasoning

See Reasoning.

string

Usage

See Usage.

Returns

this

setAttributes()

setAttributes(attributes): this

Set multiple attributes on the span at once. Warns and no-ops after end(). Mirrors OTel Span.setAttributes (and the Python SDK’s set_attributes).

Parameters

Attributes

required

Returns

this

Example

span.setAttributes({'weave.tag': 'prod', 'gen_ai.response.id': id});

Inherited from

SpanBase.setAttributes

startSubagent()

startSubagent(opts): SubAgent

Start a child SubAgent span nested under this LLM.

Parameters

SubAgentInit

required

See SubAgentInit.

Returns

SubAgent

startTool()

startTool(opts): Tool

Start a child Tool span nested under this LLM.

Parameters

ToolInit

required

See ToolInit.

Returns

Tool

think()

think(content): this

Set or extend the model’s reasoning/chain-of-thought content. Accumulates into this.reasoning.content. Folded into the last assistant message as a ReasoningPart at serialization time, matching the Python SDK’s on-the-wire shape.

Parameters

string

required

Returns

this

​Examples

​Extends

​Properties

​inputMessages

​model

​outputMessages

​providerName

​reasoning

​usage

​Methods

​addEvent()

​Parameters

​Returns

​Example

​Inherited from

​attachMedia()

​Parameters

​Returns

​attachMediaUrl()

​Parameters

​Returns

​end()

​Parameters

​Returns

​output()

​Parameters

​Returns

​record()

​Parameters

​Returns

​setAttributes()

​Parameters

​Returns

​Example

​Inherited from

​startSubagent()

​Parameters

​Returns

​startTool()

​Parameters

​Returns

​think()

​Parameters

​Returns

Examples

Extends

Properties

inputMessages

model

outputMessages

providerName

reasoning

usage

Methods

addEvent()

Parameters

Returns

Example

Inherited from

attachMedia()

Parameters

Returns

attachMediaUrl()

Parameters

Returns

end()

Parameters

Returns

output()

Parameters

Returns

record()

Parameters

Returns

setAttributes()

Parameters

Returns

Example

Inherited from

startSubagent()

Parameters

Returns

startTool()

Parameters

Returns

think()

Parameters

Returns