hsa-app/internal/classify/classify.go

// Package classify calls the Anthropic API to read an HSA receipt image and
// suggest its patient, category, date, and amount. It builds the prompt from the
// configured people and categories (see prompt.go) and forces a structured answer
// via tool-use so the result is always valid JSON. The HTTP endpoint is injectable
// so tests can run against a mock (or the cheapest model) instead of paying for Opus.
package classify

import (
	"bytes"
	"context"
	"encoding/base64"
	"encoding/json"
	"fmt"
	"io"
	"net/http"
	"strings"
	"time"

	"maisym.com/hsa/internal/config"
)

// DefaultEndpoint is the live Anthropic messages API.
const DefaultEndpoint = "https://api.anthropic.com/v1/messages"

// Suggestion is the classifier's reading of a receipt. The normalized fields are
// nil when the model could not determine them; Category falls back to the last
// (most general) configured category rather than nil, since it is a closed set.
// Raw* hold the literal text the model saw, for auditing misreads.
type Suggestion struct {
	Person   *string `json:"person"`   // canonical person label, or nil
	Category string  `json:"category"` // one of the configured category names
	Date     *string `json:"date"`     // YYYY-MM-DD, or nil
	Amount   *string `json:"amount"`   // plain number string e.g. "42.50", or nil

	RawName   string `json:"raw_name"`
	RawDate   string `json:"raw_date"`
	RawAmount string `json:"raw_amount"`

	// Usage is the token accounting for this call, used to compute its cost.
	Usage Usage
}

// Classifier holds everything needed to classify a receipt.
type Classifier struct {
	APIKey     string
	Model      string
	Endpoint   string
	HTTP       *http.Client
	Persons    []config.Person
	Categories []config.Category
}

// New builds a Classifier for the live API. Persons and categories come from the
// catalog; pass an empty model to use a sensible default.
func New(apiKey, model string, persons []config.Person, categories []config.Category) *Classifier {
	if model == "" {
		model = config.DefaultClassifyModel
	}
	return &Classifier{
		APIKey:     apiKey,
		Model:      model,
		Endpoint:   DefaultEndpoint,
		HTTP:       &http.Client{Timeout: 60 * time.Second},
		Persons:    persons,
		Categories: categories,
	}
}

// PromptPreview renders the exact system prompt that would be sent today with the
// given notes, for the read-only view on the AI tab.
func (c *Classifier) PromptPreview(today time.Time, notes []string) string {
	return BuildSystemPrompt(c.Persons, c.Categories, today.Format("2006-01-02"), notes)
}

const toolName = "record_receipt"

// --- request / response shapes ---

type apiRequest struct {
	Model       string       `json:"model"`
	MaxTokens   int          `json:"max_tokens"`
	Temperature float64      `json:"temperature"`
	System      string       `json:"system"`
	Tools       []apiTool    `json:"tools"`
	ToolChoice  apiToolPick  `json:"tool_choice"`
	Messages    []apiMessage `json:"messages"`
}

type apiTool struct {
	Name        string         `json:"name"`
	Description string         `json:"description"`
	InputSchema map[string]any `json:"input_schema"`
}

type apiToolPick struct {
	Type string `json:"type"`
	Name string `json:"name"`
}

type apiMessage struct {
	Role    string     `json:"role"`
	Content []apiBlock `json:"content"`
}

type apiBlock struct {
	Type   string     `json:"type"`
	Text   string     `json:"text,omitempty"`
	Source *apiSource `json:"source,omitempty"`
}

type apiSource struct {
	Type      string `json:"type"`
	MediaType string `json:"media_type"`
	Data      string `json:"data"`
}

type apiResponse struct {
	Content []struct {
		Type  string          `json:"type"`
		Name  string          `json:"name"`
		Input json.RawMessage `json:"input"`
	} `json:"content"`
	Usage *struct {
		InputTokens         int `json:"input_tokens"`
		OutputTokens        int `json:"output_tokens"`
		CacheReadTokens     int `json:"cache_read_input_tokens"`
		CacheCreationTokens int `json:"cache_creation_input_tokens"`
	} `json:"usage"`
	Error *struct {
		Type    string `json:"type"`
		Message string `json:"message"`
	} `json:"error"`
}

type toolInput struct {
	Person    *string `json:"person"`
	Category  *string `json:"category"`
	Date      *string `json:"date"`
	Amount    *string `json:"amount"`
	RawName   string  `json:"raw_name"`
	RawDate   string  `json:"raw_date"`
	RawAmount string  `json:"raw_amount"`
}

// Classify sends the receipt to the model and returns its normalized reading.
// today is used so the date heuristics ("closest to today, never future") are
// testable; pass time.Now().
func (c *Classifier) Classify(ctx context.Context, today time.Time, image []byte, mimeType string, notes []string) (Suggestion, error) {
	system := BuildSystemPrompt(c.Persons, c.Categories, today.Format("2006-01-02"), notes)

	imgBlock := apiBlock{
		Type:   "image",
		Source: &apiSource{Type: "base64", MediaType: mimeType, Data: base64.StdEncoding.EncodeToString(image)},
	}
	if mimeType == "application/pdf" {
		imgBlock.Type = "document"
	}

	reqBody := apiRequest{
		Model:       c.Model,
		MaxTokens:   1024,
		Temperature: 0,
		System:      system,
		Tools:       []apiTool{{Name: toolName, Description: "Record the extracted receipt fields.", InputSchema: c.inputSchema()}},
		ToolChoice:  apiToolPick{Type: "tool", Name: toolName},
		Messages: []apiMessage{{
			Role: "user",
			Content: []apiBlock{
				imgBlock,
				{Type: "text", Text: "Read this receipt and call record_receipt."},
			},
		}},
	}

	payload, err := json.Marshal(reqBody)
	if err != nil {
		return Suggestion{}, fmt.Errorf("marshal request: %w", err)
	}

	endpoint := c.Endpoint
	if endpoint == "" {
		endpoint = DefaultEndpoint
	}
	req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, bytes.NewReader(payload))
	if err != nil {
		return Suggestion{}, fmt.Errorf("build request: %w", err)
	}
	req.Header.Set("x-api-key", c.APIKey)
	req.Header.Set("anthropic-version", "2023-06-01")
	req.Header.Set("content-type", "application/json")

	httpClient := c.HTTP
	if httpClient == nil {
		httpClient = http.DefaultClient
	}
	resp, err := httpClient.Do(req)
	if err != nil {
		return Suggestion{}, fmt.Errorf("call anthropic: %w", err)
	}
	defer resp.Body.Close()

	body, err := io.ReadAll(io.LimitReader(resp.Body, 1<<20))
	if err != nil {
		return Suggestion{}, fmt.Errorf("read response: %w", err)
	}

	var ar apiResponse
	if err := json.Unmarshal(body, &ar); err != nil {
		return Suggestion{}, fmt.Errorf("parse response (status %d): %w", resp.StatusCode, err)
	}
	if ar.Error != nil {
		return Suggestion{}, fmt.Errorf("anthropic error: %s: %s", ar.Error.Type, ar.Error.Message)
	}
	if resp.StatusCode != http.StatusOK {
		return Suggestion{}, fmt.Errorf("anthropic status %d: %s", resp.StatusCode, string(body))
	}

	for _, blk := range ar.Content {
		if blk.Type == "tool_use" && blk.Name == toolName {
			var in toolInput
			if err := json.Unmarshal(blk.Input, &in); err != nil {
				return Suggestion{}, fmt.Errorf("parse tool input: %w", err)
			}
			sug := c.normalize(in)
			if ar.Usage != nil {
				sug.Usage = Usage{
					InputTokens:         ar.Usage.InputTokens,
					OutputTokens:        ar.Usage.OutputTokens,
					CacheReadTokens:     ar.Usage.CacheReadTokens,
					CacheCreationTokens: ar.Usage.CacheCreationTokens,
				}
			}
			return sug, nil
		}
	}
	return Suggestion{}, fmt.Errorf("no %s tool_use in response", toolName)
}

// inputSchema is the JSON Schema the model must fill. Enums constrain person and
// category to the configured sets; null is allowed where a field may be unknown.
func (c *Classifier) inputSchema() map[string]any {
	personEnum := append(allowedLabelsAny(c.Persons), nil)
	catEnum := make([]any, 0, len(c.Categories))
	for _, cat := range c.Categories {
		catEnum = append(catEnum, cat.Name)
	}
	return map[string]any{
		"type": "object",
		"properties": map[string]any{
			"person":     map[string]any{"type": []string{"string", "null"}, "enum": personEnum, "description": "Exact patient name, or null if absent/ambiguous."},
			"category":   map[string]any{"type": "string", "enum": catEnum, "description": "Best-fit category name."},
			"date":       map[string]any{"type": []string{"string", "null"}, "description": "Service date as YYYY-MM-DD, or null."},
			"amount":     map[string]any{"type": []string{"string", "null"}, "description": "Total paid as a plain number, or null."},
			"raw_name":   map[string]any{"type": "string", "description": "Literal name text read, or empty string."},
			"raw_date":   map[string]any{"type": "string", "description": "Literal date text read, or empty string."},
			"raw_amount": map[string]any{"type": "string", "description": "Literal amount text read, or empty string."},
		},
		"required": []string{"person", "category", "date", "amount", "raw_name", "raw_date", "raw_amount"},
	}
}

// allowedLabels here returns []any for the schema enum (string labels).
func allowedLabelsAny(persons []config.Person) []any {
	labels := allowedLabels(persons)
	out := make([]any, len(labels))
	for i, l := range labels {
		out[i] = l
	}
	return out
}

// normalize validates and cleans the raw tool output into a Suggestion. It guards
// against the model returning a non-canonical person or category despite the enum.
func (c *Classifier) normalize(in toolInput) Suggestion {
	s := Suggestion{
		RawName:   strings.TrimSpace(in.RawName),
		RawDate:   strings.TrimSpace(in.RawDate),
		RawAmount: strings.TrimSpace(in.RawAmount),
	}

	// Person: keep only if it matches a canonical label exactly.
	if in.Person != nil {
		if p := strings.TrimSpace(*in.Person); p != "" && !isNullish(p) {
			for _, want := range c.Persons {
				if strings.EqualFold(p, want.Label()) {
					label := want.Label()
					s.Person = &label
					break
				}
			}
		}
	}

	// Category: must be a configured name; otherwise fall back to the last
	// (most general) category.
	s.Category = c.fallbackCategory()
	if in.Category != nil {
		got := strings.TrimSpace(*in.Category)
		for _, cat := range c.Categories {
			if strings.EqualFold(got, cat.Name) {
				s.Category = cat.Name
				break
			}
		}
	}

	if in.Date != nil {
		if d := strings.TrimSpace(*in.Date); d != "" && !isNullish(d) {
			s.Date = &d
		}
	}
	if in.Amount != nil {
		if a := strings.TrimSpace(*in.Amount); a != "" && !isNullish(a) {
			s.Amount = &a
		}
	}
	return s
}

func (c *Classifier) fallbackCategory() string {
	if len(c.Categories) == 0 {
		return ""
	}
	return c.Categories[len(c.Categories)-1].Name
}

// isNullish catches stringified nulls the model may emit despite the schema.
func isNullish(s string) bool {
	switch strings.ToLower(s) {
	case "null", "none", "n/a", "na", "could not fetch", "unknown":
		return true
	}
	return false
}