golangLAKEHOUSE/cmd/embedd/main.go

// embedd is the embedding service. Turns text into vectors via a
// pluggable Provider (G2: Ollama at :11434). Vectors flow through
// the rest of the stack as float32 — see internal/embed for the
// boundary conversion. Default model is config-resolved; callers
// can override per request.
package main

import (
	"context"
	"encoding/json"
	"errors"
	"flag"
	"log/slog"
	"net/http"
	"os"
	"strings"
	"time"

	"github.com/go-chi/chi/v5"

	"git.agentview.dev/profit/golangLAKEHOUSE/internal/embed"
	"git.agentview.dev/profit/golangLAKEHOUSE/internal/shared"
)

const (
	maxRequestBytes = 4 << 20         // 4 MiB cap on /embed body — texts plural
	batchDeadline   = 60 * time.Second // upper bound on a single /embed batch
)

func main() {
	configPath := flag.String("config", "lakehouse.toml", "path to TOML config")
	flag.Parse()

	cfg, err := shared.LoadConfig(*configPath)
	if err != nil {
		slog.Error("config", "err", err)
		os.Exit(1)
	}
	if cfg.Embedd.ProviderURL == "" {
		slog.Error("config", "err", "embedd.provider_url is required")
		os.Exit(1)
	}

	// Wrap the upstream provider in an LRU cache so repeat queries
	// (the staffing co-pilot replays many of the same texts) bypass
	// the ~50ms Ollama round-trip. Cache size 0 = pass-through.
	base := embed.NewOllama(cfg.Embedd.ProviderURL, cfg.Embedd.DefaultModel)
	cached, err := embed.NewCachedProvider(base, cfg.Embedd.DefaultModel, cfg.Embedd.CacheSize)
	if err != nil {
		slog.Error("embed cache", "err", err)
		os.Exit(1)
	}
	slog.Info("embed cache",
		"size", cfg.Embedd.CacheSize,
		"default_model", cfg.Embedd.DefaultModel,
		"enabled", cfg.Embedd.CacheSize > 0)
	h := &handlers{provider: cached, cache: cached}

	if err := shared.Run("embedd", cfg.Embedd.Bind, h.register, cfg.Auth); err != nil {
		slog.Error("server", "err", err)
		os.Exit(1)
	}
}

type handlers struct {
	provider embed.Provider
	// cache is the same instance as provider when caching is enabled,
	// kept as a typed pointer so /v1/embed/stats can expose hit-rate
	// without type-asserting through the Provider interface. nil when
	// CacheSize=0 (pass-through mode).
	cache *embed.CachedProvider
}

func (h *handlers) register(r chi.Router) {
	r.Post("/embed", h.handleEmbed)
	r.Get("/embed/stats", h.handleStats)
}

// handleStats reports cache hits/misses + hit rate + size. Operators
// use this to confirm the cache is doing its job (high hit rate) or
// is sized wrong (low hit rate + many misses on a workload that
// should have repeats).
func (h *handlers) handleStats(w http.ResponseWriter, _ *http.Request) {
	w.Header().Set("Content-Type", "application/json")
	if h.cache == nil {
		_ = json.NewEncoder(w).Encode(map[string]any{"enabled": false})
		return
	}
	hits, misses := h.cache.Stats()
	_ = json.NewEncoder(w).Encode(map[string]any{
		"enabled":   true,
		"hits":      hits,
		"misses":    misses,
		"hit_rate":  h.cache.HitRate(),
		"size":      h.cache.Len(),
	})
}

// embedRequest is the POST /embed body. Texts is the list to
// embed; Model is optional (empty → use server default).
type embedRequest struct {
	Texts []string `json:"texts"`
	Model string   `json:"model,omitempty"`
}

func (h *handlers) handleEmbed(w http.ResponseWriter, r *http.Request) {
	defer r.Body.Close()
	r.Body = http.MaxBytesReader(w, r.Body, maxRequestBytes)
	var req embedRequest
	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
		var maxErr *http.MaxBytesError
		if errors.As(err, &maxErr) || strings.Contains(err.Error(), "http: request body too large") {
			http.Error(w, "body too large", http.StatusRequestEntityTooLarge)
			return
		}
		http.Error(w, "decode body: "+err.Error(), http.StatusBadRequest)
		return
	}

	// Per scrum O-W3 (Opus): reject empty strings up front. Ollama's
	// behavior on empty prompt is version-dependent (some return
	// errors, some return zero vectors); rejecting at the boundary
	// gives callers a deterministic 400 instead of 502.
	for j, t := range req.Texts {
		if t == "" {
			http.Error(w, "texts["+itoa(j)+"]: empty string", http.StatusBadRequest)
			return
		}
	}

	// Per scrum C1 (Opus + Kimi convergent): per-text 60s timeout
	// without a batch-level cap means a 100-text batch with one
	// stuck call can pin the handler for ~6000s. Set a hard batch
	// ceiling derived from the request ctx so a wedged Ollama
	// surfaces as 504-ish (mapped to 502 by the upstream-error
	// path below) rather than holding the connection forever.
	ctx, cancel := context.WithTimeout(r.Context(), batchDeadline)
	defer cancel()

	res, err := h.provider.Embed(ctx, req.Texts, req.Model)
	if errors.Is(err, embed.ErrEmptyTexts) {
		http.Error(w, err.Error(), http.StatusBadRequest)
		return
	}
	if errors.Is(err, embed.ErrModelMismatch) {
		http.Error(w, err.Error(), http.StatusBadGateway)
		return
	}
	if err != nil {
		// Upstream-shape errors (Ollama down, model missing,
		// 5xx body) bubble up as 502 — distinguishes "your input
		// was wrong" (400) from "the embedding backend was wrong" (502).
		slog.Warn("embed", "err", err)
		http.Error(w, "embed: "+err.Error(), http.StatusBadGateway)
		return
	}
	w.Header().Set("Content-Type", "application/json")
	if err := json.NewEncoder(w).Encode(res); err != nil {
		slog.Warn("embed encode", "err", err)
	}
}

// itoa is a tiny helper for error messages without pulling strconv
// in just for one call site.
func itoa(i int) string {
	if i == 0 {
		return "0"
	}
	var buf [20]byte
	pos := len(buf)
	for i > 0 {
		pos--
		buf[pos] = byte('0' + i%10)
		i /= 10
	}
	return string(buf[pos:])
}