Documentation
¶
Overview ¶
Package llm is the contract every core-models provider adapter implements. It mirrors ADK's model package field for field, so that an adapter module (adkv2 today) can adapt it to an ADK major with plain copies — and it imports no ADK, so the core module never puts an ADK major into a consumer's build.
Each ADK major's model.LLM is identical in shape to the others and distinct as a Go type; this package is the neutral spelling every major can reach (docs/design.md §4, decision D2).
Index ¶
Constants ¶
This section is empty.
Variables ¶
var ErrEmptyResponse = errors.New("model returned no usable content")
ErrEmptyResponse is what an adapter's error wraps when the model answered with no usable content, no finish reason other than STOP, and no error: the silent turn an agent loop otherwise waits on forever (mast #220). Adapters wrap it with their own wording, so a caller can recognise "the model said nothing" with errors.Is without importing any adapter — a one-shot side question, for which an empty answer is the answer, needs exactly that (core-agent's models sentinel, generalised).
Functions ¶
This section is empty.
Types ¶
type LLM ¶
type LLM interface {
// Name reports the model this LLM serves.
Name() string
// GenerateContent sends req. With stream false it yields one
// complete Response; with stream true it yields partial Responses
// (Partial set) followed by a final complete one.
GenerateContent(ctx context.Context, req *Request, stream bool) iter.Seq2[*Response, error]
}
LLM is a model a caller can send one request to. It is ADK's model.LLM over this package's Request and Response.
type Request ¶
type Request struct {
Model string
Contents []*genai.Content
Config *genai.GenerateContentConfig
}
Request mirrors ADK's model.LLMRequest without its Tools field. That field maps tool names to ADK tool objects for ADK's own flows to dispatch on; no adapter reads it, since the declarations a provider sends come from Config.Tools.
type Response ¶
type Response struct {
Content *genai.Content `json:"content,omitempty"`
CitationMetadata *genai.CitationMetadata `json:"citationMetadata,omitempty"`
GroundingMetadata *genai.GroundingMetadata `json:"groundingMetadata,omitempty"`
UsageMetadata *genai.GenerateContentResponseUsageMetadata `json:"usageMetadata,omitempty"`
CustomMetadata map[string]any `json:"customMetadata,omitempty"`
LogprobsResult *genai.LogprobsResult `json:"logprobsResult,omitempty"`
InputTranscription *genai.Transcription `json:"inputTranscription,omitempty"`
OutputTranscription *genai.Transcription `json:"outputTranscription,omitempty"`
ModelVersion string `json:"modelVersion,omitempty"`
Partial bool `json:"partial,omitempty"`
TurnComplete bool `json:"turnComplete,omitempty"`
Interrupted bool `json:"interrupted,omitempty"`
SessionResumptionHandle string `json:"sessionResumptionHandle,omitempty"`
ErrorCode string `json:"errorCode,omitempty"`
ErrorMessage string `json:"errorMessage,omitempty"`
FinishReason genai.FinishReason `json:"finishReason,omitempty"`
AvgLogprobs float64 `json:"avgLogprobs,omitempty"`
}
Response mirrors ADK's model.LLMResponse, JSON tags included (ADK v2 spells them; v1 has none, and Go's decoder matches either).
CustomMetadata is where an adapter attaches what genai's usage struct cannot hold — see package usage.
Every dialect yields exactly one response with TurnComplete set per call: the single response of a unary call, or the aggregate that ends a stream. Streamed chunks are Partial and never TurnComplete, so a consumer that commits usage on TurnComplete counts each call once.