Affected by 
Documentation
¶
Overview ¶
Package rerank adapts rerank protocols to application batching and score semantics.
Index ¶
Constants ¶
const ( LKEAPRerankEndpoint = tencentlkeap.LKEAPRerankEndpoint LKEAPDefaultRegion = tencentlkeap.LKEAPDefaultRegion LKEAPDefaultRerankModel = tencentlkeap.LKEAPDefaultRerankModel )
Tencent action defaults retained for existing callers.
Variables ¶
var ErrImagesUnsupported = errors.New("rerank model does not accept images")
ErrImagesUnsupported is returned when a reranker is asked to score images its model or endpoint cannot take.
Functions ¶
func MaxPassageRunes ¶
MaxPassageRunes reports r's passage limit for query, or 0 when r documents none.
Types ¶
type DocumentInfo ¶
type DocumentInfo struct {
Text string `json:"text"`
}
func (*DocumentInfo) UnmarshalJSON ¶
func (d *DocumentInfo) UnmarshalJSON(data []byte) error
UnmarshalJSON handles both string and object formats for DocumentInfo
type ImageReranker ¶
type ImageReranker interface {
// AcceptsImages reports whether both the model and its endpoint take
// image documents.
AcceptsImages() bool
// ImageLimits are the vendor's documented per-image limits. The caller
// prepares an image to fit; one outside them is refused.
ImageLimits() imageprep.Limits
// RerankImages scores images against the query on the same scale as
// Rerank scores text, best first; indices refer to images.
RerankImages(ctx context.Context, query string, images []Image) ([]RankResult, error)
}
ImageReranker is the image side of a Reranker whose model scores an image against a text query by what it shows. Most rerankers have none; use AsImageReranker rather than a type assertion, since every decorator implements it whether or not the model underneath does.
func AsImageReranker ¶
func AsImageReranker(r Reranker) (ImageReranker, bool)
AsImageReranker returns r's image side when its model scores images.
type PassageLimiter ¶
type PassageLimiter interface {
// MaxPassageRunes returns the longest document, in runes, one request
// can carry beside query. 0 means no documented limit.
MaxPassageRunes(query string) int
}
PassageLimiter is implemented by rerankers whose vendor documents how large a document may be. The rerank stage builds its passages itself (title, chunk body, captions, OCR text, generated questions), so it can fit them to the limit instead of letting one oversized candidate fail the whole request, which the protocol layer rightly refuses to truncate on its own.
type RankResult ¶
type RankResult struct {
Index int `json:"index"`
Document DocumentInfo `json:"document"`
RelevanceScore float64 `json:"relevance_score"`
}
func (*RankResult) UnmarshalJSON ¶
func (r *RankResult) UnmarshalJSON(data []byte) error
Handles the RelevanceScore field by checking if RelevanceScore exists first, otherwise falls back to Score field
type Reranker ¶
type Reranker interface {
// Rerank reranks documents based on relevance to the query
Rerank(ctx context.Context, query string, documents []string) ([]RankResult, error)
// GetModelName returns the model name
GetModelName() string
// GetModelID returns the model ID
GetModelID() string
}
Reranker defines the interface for document reranking
func NewReranker ¶
func NewReranker(config *RerankerConfig) (Reranker, error)
NewReranker creates a reranker based on the configuration
type RerankerConfig ¶
type RerankerConfig struct {
APIKey string
BaseURL string
ModelName string
Source types.ModelSource
ModelID string
Provider string // Provider identifier: openai, aliyun, zhipu, siliconflow, jina, generic
Spec *types.ModelSpecOverride `json:"spec,omitempty"`
ExtraConfig map[string]string
// CustomHeaders 允许在调用远程 API 时附加自定义 HTTP 请求头（类似 OpenAI Python SDK 的 extra_headers）。
CustomHeaders map[string]string
AppID string
AppSecret string // 加密值，工厂函数调用方传入，使用前已解密
}
func ConfigFromModel ¶
func ConfigFromModel(m *types.Model, appID, appSecret string) *RerankerConfig
ConfigFromModel 根据 types.Model 构造 RerankerConfig。 生产路径（从 DB 拉起）和测试连接路径（临时表单）共享这份映射。 appID / appSecret 是已解密的 WeKnoraCloud 凭证，调用方负责传入。
Source Files