package inference

// The wire types for the MEGAMIND inference plane. They are deliberately plain JSON structs shared in
// spirit with cmd/megamind-gateway, so the typed client and the stateless service agree on one contract.
// Vectors are float32 to match pgvector(384); the API never sends client identifiers to the plane.

// EmbedRequest asks for embeddings of one or more input strings under a named model.
type EmbedRequest struct {
	Model  string   `json:"model"`
	Inputs []string `json:"inputs"`
}

// EmbedResponse returns one vector per input plus the dimension, so the caller can hard fail on a
// mismatch with the pgvector column width.
type EmbedResponse struct {
	Model      string      `json:"model"`
	Dimension  int         `json:"dimension"`
	Embeddings [][]float32 `json:"embeddings"`
}

// RerankRequest scores documents against a query with a cross encoder (or a cosine fallback).
type RerankRequest struct {
	Model     string   `json:"model"`
	Query     string   `json:"query"`
	Documents []string `json:"documents"`
	TopN      int      `json:"top_n"`
}

// RerankResult is one scored document, identified only by its index in the request (never by a client id).
type RerankResult struct {
	Index int     `json:"index"`
	Score float64 `json:"score"`
}

// RerankResponse returns the reranked results, highest score first.
type RerankResponse struct {
	Model   string         `json:"model"`
	Results []RerankResult `json:"results"`
}

// GenerateRequest is a single stateless generation. System carries instructions; Prompt the question and
// retrieved evidence. The plane keeps nothing between calls.
type GenerateRequest struct {
	Model       string  `json:"model"`
	System      string  `json:"system"`
	Prompt      string  `json:"prompt"`
	MaxTokens   int     `json:"max_tokens"`
	Temperature float64 `json:"temperature"`
}

// GenerateResponse is the model's text plus light accounting.
type GenerateResponse struct {
	Model  string `json:"model"`
	Text   string `json:"text"`
	Tokens int    `json:"tokens"`
}

// FaithfulnessRequest asks whether Answer is supported by Evidence. It is the self correction gate's input.
type FaithfulnessRequest struct {
	Answer   string   `json:"answer"`
	Evidence []string `json:"evidence"`
}

// FaithfulnessResponse reports support as a 0 to 1 score plus a boolean decision and a short reason.
type FaithfulnessResponse struct {
	Supported bool    `json:"supported"`
	Score     float64 `json:"score"`
	Reason    string  `json:"reason"`
}