embedding db

Makefile

@@ -1,6 +1,6 @@
 # Makefile for running the Vet Clinic Chat Assistant locally with Ollama
 
-.PHONY: run ollama-start ollama-stop ollama-pull ollama-status
+.PHONY: run ollama-start ollama-stop ollama-pull ollama-status curl-embed curl-translate curl-chat
 
 # Start Ollama server (if not already running)
 ollama-start:
@@ -20,6 +20,15 @@ ollama-pull:
 ollama-status:
 	ollama list
 
+# Ollama host & models (override as needed)
+OLLAMA_HOST ?= http://localhost:11434
+# Primary chat / reasoning model (already using OPENAI_MODEL var for compatibility)
+OPENAI_MODEL ?= qwen3:latest
+# Optional separate embedding model
+OLLAMA_EMBED_MODEL ?= all-minilm
+# Translation prompt (mirrors config.yaml translate_prompt). Can override: make curl-translate PROMPT="..." TRANSLATE_PROMPT="..."
+TRANSLATE_PROMPT ?= Translate the following veterinary-related sentence to English. Input: '$(PROMPT)'. Return ONLY the English translation, no extra text, no markdown, no quotes. If already English, return as is.
+
 # Database configuration (override via: make run DB_PASSWORD=secret DB_NAME=other)
 DB_HOST ?= localhost
 DB_PORT ?= 5432
@@ -70,3 +79,31 @@ test-coverage:
 # Run tests with HTML coverage report
 test-coverage-html: test-coverage
 	go tool cover -html=coverage.out
+
+# --- Utility curl targets ---
+# Example: make curl-embed PROMPT="warm up"
+curl-embed:
+	@test -n "$(PROMPT)" || { echo "Usage: make curl-embed PROMPT='text' [OLLAMA_EMBED_MODEL=model]"; exit 1; }
+	@echo "[curl-embed] model=$(OLLAMA_EMBED_MODEL) prompt='$(PROMPT)'"
+	@curl -sS -X POST "$(OLLAMA_HOST)/api/embeddings" \
+	  -H 'Content-Type: application/json' \
+	  -d '{"model":"$(OLLAMA_EMBED_MODEL)","prompt":"$(PROMPT)"}' | jq . || true
+
+# Example: make curl-translate PROMPT="A kutyám nem eszik"
+curl-translate:
+	@test -n "$(PROMPT)" || { echo "Usage: make curl-translate PROMPT='sentence to translate'"; exit 1; }
+	@echo "[curl-translate] model=$(OPENAI_MODEL)"; \
+	PROMPT_JSON=$$(printf '%s' "$(TRANSLATE_PROMPT)" | jq -Rs .); \
+	curl -sS -X POST "$(OLLAMA_HOST)/api/chat" \
+	  -H 'Content-Type: application/json' \
+	  -d '{"model":"$(OPENAI_MODEL)","messages":[{"role":"user","content":'$$PROMPT_JSON'}],"stream":false}' | jq -r '.message.content' || true
+
+# Generic chat invocation (raw user PROMPT)
+# Example: make curl-chat PROMPT="List 3 dog breeds"
+curl-chat:
+	@test -n "$(PROMPT)" || { echo "Usage: make curl-chat PROMPT='your message'"; exit 1; }
+	@echo "[curl-chat] model=$(OPENAI_MODEL)"; \
+	PROMPT_JSON=$$(printf '%s' "$(PROMPT)" | jq -Rs .); \
+	curl -sS -X POST "$(OLLAMA_HOST)/api/chat" \
+	  -H 'Content-Type: application/json' \
+	  -d '{"model":"$(OPENAI_MODEL)","messages":[{"role":"user","content":'$$PROMPT_JSON'}],"stream":false}' | jq -r '.message.content' || true

config/config.yaml

@@ -1,4 +1,4 @@
 llm:
   extract_keywords_prompt: "You will extract structured data from the user input. Input text: {{.Message}}. Return ONLY valid minified JSON object with keys: translate (English translation of input), keyword (array of 3-5 concise English veterinary-related keywords derived strictly from the input), animal (animal mentioned or 'unknown'). Example: {\"translate\":\"dog has diarrhea\",\"keyword\":[\"diarrhea\",\"digestive\"],\"animal\":\"dog\"}. Do not add extra text, markdown, or quotes outside JSON."
   disambiguate_prompt: "Given candidate visit entries (JSON array): {{.Entries}} and user message: {{.Message}} choose the best matching visit's ID. Return ONLY JSON: {\"visitReason\":\"<one of the candidate IDs or empty string if none>\"}. No other text."
-  translate_prompt: "Translate the following veterinary-related sentence to English. Input: '{{.Message}}'. Return ONLY the English translation, no extra text, no markdown, no quotes. If already English, return as is."
+  translate_prompt: "Translate the following veterinary-related sentence to English. Input: '{{.Message}}'. Return ONLY the English translation as one concise sentence. IMPORTANT: Do NOT output any <think> tags, reasoning, analysis, or explanations. No markdown, no quotes. If already English, return it unchanged."

llm.go

@@ -55,7 +55,14 @@ func NewLLMClientFromEnv(repo ChatRepositoryAPI) LLMClientAPI {
 	case "openai", "openrouter":
 		return NewOpenAIClient(apiKey, baseURL, model, repo)
 	case "ollama", "":
-		return NewOllamaClient(apiKey, baseURL, model, repo)
+		oc := NewOllamaClient(apiKey, baseURL, model, repo)
+		em := os.Getenv("OLLAMA_EMBED_MODEL")
+		if strings.TrimSpace(em) == "" {
+			em = "all-minilm"
+			logrus.Infof("No OLLAMA_EMBED_MODEL specified; defaulting embedding model to %s", em)
+		}
+		oc.EmbeddingModel = em
+		return oc
 	default:
 		logrus.Warnf("Unknown LLM_PROVIDER %q, defaulting to Ollama", provider)
 		return NewOllamaClient(apiKey, baseURL, model, repo)

main.go

@@ -72,6 +72,9 @@ func main() {
 	// Initialize LLM client
 	llm := NewLLMClientFromEnv(repo)
 
+	// Launch background backfill of sentence embeddings (non-blocking)
+	startSentenceEmbeddingBackfill(repo, llm, &visitDB)
+
 	// Wrap templates for controller
 	uiTmpl := &TemplateWrapper{Tmpl: uiTemplate}
 	uiDBEditTmpl := &TemplateWrapper{Tmpl: uiDBEditTemplate}

migrations/0004_alter_sentence_embeddings_visit_id_type.up.sql

@@ -0,0 +1,8 @@
+-- +goose Up
+ALTER TABLE sentence_embeddings
+    ALTER COLUMN visit_id TYPE TEXT USING visit_id::text;
+
+-- +goose Down
+ALTER TABLE sentence_embeddings
+    ALTER COLUMN visit_id TYPE INTEGER USING visit_id::integer;
+

migrations/0005_add_unique_idx_sentence_embeddings.up.sql

@@ -0,0 +1,6 @@
+-- +goose Up
+CREATE UNIQUE INDEX IF NOT EXISTS ux_sentence_embeddings_visit_sentence
+    ON sentence_embeddings(visit_id, sentence);
+
+-- +goose Down
+DROP INDEX IF EXISTS ux_sentence_embeddings_visit_sentence;

migrations/0006_alter_sentence_embeddings_dim_384.up.sql

@@ -0,0 +1,14 @@
+-- +goose Up
+-- WARNING: This alters the embeddings vector dimension from 1536 to 384.
+-- Ensure you are switching to a 384-dim embedding model (e.g., all-minilm).
+-- If existing rows have 1536-d vectors this command will fail; you may need to
+-- TRUNCATE TABLE sentence_embeddings first (after backing up) before applying.
+ALTER TABLE sentence_embeddings
+    ALTER COLUMN embeddings TYPE vector(384);
+
+-- +goose Down
+-- Revert to 1536 dimensions (for models like OpenAI text-embedding-3-large).
+-- Will fail if existing rows are 384-d.
+ALTER TABLE sentence_embeddings
+    ALTER COLUMN embeddings TYPE vector(1536);
+

migrations/0007_add_dual_embedding_columns.up.sql

@@ -0,0 +1,13 @@
+-- +goose Up
+-- Add separate columns for different embedding dimensions.
+-- Existing 'embeddings' column (if present) is left untouched for backward compatibility.
+-- Application code will now populate embedding_384 or embedding_1536 instead.
+ALTER TABLE sentence_embeddings
+    ADD COLUMN IF NOT EXISTS embedding_384   vector(384),
+    ADD COLUMN IF NOT EXISTS embedding_1536  vector(1536);
+
+-- +goose Down
+ALTER TABLE sentence_embeddings
+    DROP COLUMN IF EXISTS embedding_384,
+    DROP COLUMN IF EXISTS embedding_1536;
+

migrations/0008_drop_legacy_embeddings_column.up.sql

@@ -0,0 +1,11 @@
+-- +goose Up
+-- Drop the legacy single-dimension embeddings column (was NOT NULL) to allow inserts
+-- that now use embedding_384 / embedding_1536. All new data goes into those columns.
+ALTER TABLE sentence_embeddings
+    DROP COLUMN IF EXISTS embeddings;
+
+-- +goose Down
+-- Re-create the legacy embeddings column (empty) as vector(1536) NULLABLE for rollback.
+ALTER TABLE sentence_embeddings
+    ADD COLUMN IF NOT EXISTS embeddings vector(1536);
+

ollama_client.go

@@ -7,7 +7,10 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"os"
+	"strconv"
 	"strings"
+	"time"
 
 	"github.com/sirupsen/logrus"
 )
@@ -15,10 +18,11 @@ import (
 // --- OllamaClient implementation ---
 
 type OllamaClient struct {
-	APIKey  string
+	APIKey         string
-	BaseURL string
+	BaseURL        string
-	Model   string
+	Model          string
-	Repo    ChatRepositoryAPI
+	EmbeddingModel string
+	Repo           ChatRepositoryAPI
 }
 
 func NewOllamaClient(apiKey, baseURL, model string, repo ChatRepositoryAPI) *OllamaClient {
@@ -90,12 +94,28 @@ func (llm *OllamaClient) ollamaCompletion(ctx context.Context, prompt string, fo
 	if apiURL == "" {
 		apiURL = "http://localhost:11434/api/chat"
 	}
+
+	messages := []map[string]string{{"role": "user", "content": prompt}}
+	//if os.Getenv("DISABLE_THINK") == "1" {
+	// System message to suppress chain-of-thought style outputs.
+	messages = append([]map[string]string{{
+		"role":    "system",
+		"content": "You are a concise assistant. Output ONLY the final answer requested by the user. Do not include reasoning, analysis, or <think> tags.",
+	}}, messages...)
+	//}
+
 	body := map[string]interface{}{
 		"model":    llm.Model,
-		"messages": []map[string]string{{"role": "user", "content": prompt}},
+		"messages": messages,
 		"stream":   false,
 		"format":   format,
 	}
+
+	// Optional: Add a stop sequence to prevent <think> tags if they appear
+	if os.Getenv("DISABLE_THINK") == "1" {
+		body["options"] = map[string]interface{}{"stop": []string{"<think>"}}
+	}
+
 	jsonBody, _ := json.Marshal(body)
 	req, _ := http.NewRequestWithContext(ctx, http.MethodPost, apiURL, bytes.NewBuffer(jsonBody))
 	if llm.APIKey != "" {
@@ -128,43 +148,112 @@ func (llm *OllamaClient) ollamaCompletion(ctx context.Context, prompt string, fo
 	return "", fmt.Errorf("unrecognized LLM response format: %.200s", string(raw))
 }
 
+func normalizeOllamaHost(raw string) string {
+	if raw == "" {
+		return "http://localhost:11434"
+	}
+	// strip trailing /api/* paths if user provided full endpoint
+	lower := strings.ToLower(raw)
+	for _, seg := range []string{"/api/chat", "/api/embeddings", "/api/generate"} {
+		if strings.HasSuffix(lower, seg) {
+			return raw[:len(raw)-len(seg)]
+		}
+	}
+	return raw
+}
+
 func (llm *OllamaClient) GetEmbeddings(ctx context.Context, input string) ([]float64, error) {
-	apiURL := llm.BaseURL
+	host := normalizeOllamaHost(llm.BaseURL)
-	if apiURL == "" {
+	apiURL := host + "/api/embeddings"
-		apiURL = "http://localhost:11434/api/embeddings"
+	modelName := llm.Model
+	if llm.EmbeddingModel != "" {
+		modelName = llm.EmbeddingModel
 	}
-	body := map[string]interface{}{
+	// retry parameters (env override OLLAMA_EMBED_ATTEMPTS)
-		"model":  llm.Model,
+	maxAttempts := 5
-		"prompt": input,
+	if v := os.Getenv("OLLAMA_EMBED_ATTEMPTS"); v != "" {
+		if n, err := strconv.Atoi(v); err == nil && n > 0 && n < 20 {
+			maxAttempts = n
+		}
 	}
-	jsonBody, _ := json.Marshal(body)
+	baseBackoff := 300 * time.Millisecond
-	req, _ := http.NewRequestWithContext(ctx, http.MethodPost, apiURL, bytes.NewBuffer(jsonBody))
+	var lastErr error
-	if llm.APIKey != "" {
+	for attempt := 0; attempt < maxAttempts; attempt++ {
-		req.Header.Set("Authorization", "Bearer "+llm.APIKey)
+		select {
+		case <-ctx.Done():
+			return nil, ctx.Err()
+		default:
+		}
+		body := map[string]interface{}{
+			"model":  modelName,
+			"prompt": input,
+		}
+		jsonBody, _ := json.Marshal(body)
+		req, _ := http.NewRequestWithContext(ctx, http.MethodPost, apiURL, bytes.NewBuffer(jsonBody))
+		if llm.APIKey != "" {
+			req.Header.Set("Authorization", "Bearer "+llm.APIKey)
+		}
+		req.Header.Set("Content-Type", "application/json")
+		req.Header.Set("Accept", "application/json")
+		resp, err := (&http.Client{}).Do(req)
+		if err != nil {
+			lastErr = err
+			logrus.WithError(err).Warnf("[Ollama] embeddings request attempt=%d failed", attempt+1)
+		} else {
+			raw, rerr := io.ReadAll(resp.Body)
+			resp.Body.Close()
+			if rerr != nil {
+				lastErr = rerr
+			} else {
+				var generic map[string]json.RawMessage
+				if jerr := json.Unmarshal(raw, &generic); jerr != nil {
+					lastErr = fmt.Errorf("unrecognized response (parse): %w", jerr)
+				} else if embRaw, ok := generic["embedding"]; ok && len(embRaw) > 0 {
+					var emb []float64
+					if jerr := json.Unmarshal(embRaw, &emb); jerr != nil {
+						lastErr = fmt.Errorf("failed to decode embedding: %w", jerr)
+					} else if len(emb) == 0 {
+						lastErr = fmt.Errorf("empty embedding returned")
+					} else {
+						return emb, nil
+					}
+				} else if drRaw, ok := generic["done_reason"]; ok {
+					var reason string
+					_ = json.Unmarshal(drRaw, &reason)
+					if reason == "load" { // transient model loading state
+						lastErr = fmt.Errorf("model loading")
+					} else {
+						lastErr = fmt.Errorf("unexpected done_reason=%s", reason)
+					}
+				} else if errRaw, ok := generic["error"]; ok {
+					var errMsg string
+					_ = json.Unmarshal(errRaw, &errMsg)
+					if errMsg != "" {
+						lastErr = fmt.Errorf("embedding error: %s", errMsg)
+					} else {
+						lastErr = fmt.Errorf("embedding error (empty message)")
+					}
+				} else {
+					lastErr = fmt.Errorf("unrecognized embedding response: %.200s", string(raw))
+				}
+			}
+		}
+		if lastErr == nil {
+			break
+		}
+		// backoff if not last attempt
+		if attempt < maxAttempts-1 {
+			delay := baseBackoff << attempt
+			if strings.Contains(strings.ToLower(lastErr.Error()), "model loading") {
+				delay += 1 * time.Second
+			}
+			time.Sleep(delay)
+		}
 	}
-	req.Header.Set("Content-Type", "application/json")
+	if lastErr == nil {
-	req.Header.Set("Accept", "application/json")
+		lastErr = fmt.Errorf("embedding retrieval failed with no error info")
-	client := &http.Client{}
-	resp, err := client.Do(req)
-	if err != nil {
-		return nil, err
 	}
-	defer resp.Body.Close()
+	return nil, lastErr
-	raw, err := io.ReadAll(resp.Body)
-	if err != nil {
-		return nil, err
-	}
-	var ollama struct {
-		Embedding []float64 `json:"embedding"`
-		Error     string    `json:"error"`
-	}
-	if err := json.Unmarshal(raw, &ollama); err == nil && len(ollama.Embedding) > 0 {
-		return ollama.Embedding, nil
-	}
-	if ollama.Error != "" {
-		return nil, fmt.Errorf("embedding error: %s", ollama.Error)
-	}
-	return nil, fmt.Errorf("unrecognized embedding response: %.200s", string(raw))
 }
 
 func (llm *OllamaClient) TranslateToEnglish(ctx context.Context, message string) (string, error) {

repository.go

@@ -2,7 +2,10 @@ package main
 
 import (
 	"context"
+	"fmt"
 	"os"
+	"strconv"
+	"strings"
 	"time"
 
 	"github.com/jackc/pgx/v5/pgconn"
@@ -263,6 +266,69 @@ func (r *PGChatRepository) CreateUser(ctx context.Context, username, passwordHas
 	return err
 }
 
+// InsertSentenceEmbedding inserts a sentence embedding if not already present (unique index on visit_id,sentence)
+func (r *PGChatRepository) InsertSentenceEmbedding(ctx context.Context, visitID, sentence, translated string, embedding []float64) error {
+	if r == nil || r.pool == nil {
+		return nil
+	}
+	l := len(embedding)
+	if l != 384 && l != 1536 {
+		err := fmt.Errorf("unsupported embedding length %d (expected 384 or 1536)", l)
+		logrus.WithError(err).Warn("skipping sentence embedding insert")
+		return err
+	}
+	// Build vector literal
+	var b strings.Builder
+	b.Grow(len(embedding)*8 + 2)
+	b.WriteByte('[')
+	for i, v := range embedding {
+		if i > 0 {
+			b.WriteByte(',')
+		}
+		b.WriteString(strconv.FormatFloat(v, 'f', -1, 64))
+	}
+	b.WriteByte(']')
+	vecLiteral := b.String()
+
+	ctx, cancel := context.WithTimeout(ctx, 6*time.Second)
+	defer cancel()
+
+	var sqlStmt string
+	if l == 384 {
+		sqlStmt = `INSERT INTO sentence_embeddings (visit_id, sentence, translated, embedding_384)
+			VALUES ($1,$2,$3,$4::vector)
+			ON CONFLICT (visit_id, sentence) DO UPDATE
+			SET embedding_384 = EXCLUDED.embedding_384,
+			    translated = COALESCE(sentence_embeddings.translated, EXCLUDED.translated)`
+	} else { // 1536
+		sqlStmt = `INSERT INTO sentence_embeddings (visit_id, sentence, translated, embedding_1536)
+			VALUES ($1,$2,$3,$4::vector)
+			ON CONFLICT (visit_id, sentence) DO UPDATE
+			SET embedding_1536 = EXCLUDED.embedding_1536,
+			    translated = COALESCE(sentence_embeddings.translated, EXCLUDED.translated)`
+	}
+	_, err := r.pool.Exec(ctx, sqlStmt, visitID, sentence, translated, vecLiteral)
+	if err != nil {
+		logrus.WithError(err).Warn("failed to upsert sentence embedding (dual columns)")
+	}
+	return err
+}
+
+// ExistsSentenceEmbedding checks if a sentence embedding exists
+func (r *PGChatRepository) ExistsSentenceEmbedding(ctx context.Context, visitID, sentence string) (bool, error) {
+	if r == nil || r.pool == nil {
+		return false, nil
+	}
+	ctx, cancel := context.WithTimeout(ctx, 2*time.Second)
+	defer cancel()
+	var exists bool
+	err := r.pool.QueryRow(ctx, `SELECT EXISTS (SELECT 1 FROM sentence_embeddings WHERE visit_id=$1 AND sentence=$2)`, visitID, sentence).Scan(&exists)
+	if err != nil {
+		return false, err
+	}
+	return exists, nil
+}
+
 // Close releases pool resources
 func (r *PGChatRepository) Close() {
 	if r != nil && r.pool != nil {

sentence_embeddings.go

@@ -0,0 +1,154 @@
+package main
+
+import (
+	"context"
+	"os"
+	"regexp"
+	"strings"
+	"time"
+
+	"github.com/sirupsen/logrus"
+)
+
+var sentenceSplitRegex = regexp.MustCompile(`(?m)(?:[^.!?\n]+[.!?]|[^.!?\n]+$)`)
+
+// configurable via env (seconds); defaults chosen for model cold start friendliness
+func envDuration(key string, def time.Duration) time.Duration {
+	if v := os.Getenv(key); v != "" {
+		if d, err := time.ParseDuration(v); err == nil {
+			return d
+		}
+	}
+	return def
+}
+
+// startSentenceEmbeddingBackfill launches a background goroutine that iterates all visits
+// and stores (visit_id, sentence, translated, embedding) records in sentence_embeddings table
+// if they do not already exist (relying on unique index ON CONFLICT DO NOTHING).
+func startSentenceEmbeddingBackfill(repo *PGChatRepository, llm LLMClientAPI, vdb *VisitDB) {
+	if repo == nil || llm == nil || vdb == nil {
+		logrus.Info("Sentence embedding backfill skipped (missing repo, llm or vdb)")
+		return
+	}
+	if disable := strings.ToLower(os.Getenv("SENTENCE_BACKFILL_DISABLE")); disable == "1" || disable == "true" {
+		logrus.Info("Sentence embedding backfill disabled via SENTENCE_BACKFILL_DISABLE env var")
+		return
+	}
+	translateTimeout := envDuration("TRANSLATE_TIMEOUT", 45*time.Second)
+	embeddingTimeout := envDuration("EMBEDDING_TIMEOUT", 45*time.Second)
+	maxTranslateAttempts := 3
+	maxEmbeddingAttempts := 3
+	go func() {
+		start := time.Now()
+		logrus.WithFields(logrus.Fields{"translateTimeout": translateTimeout, "embeddingTimeout": embeddingTimeout}).Info("Sentence embedding backfill started")
+		processed := 0
+		inserted := 0
+		skippedExisting := 0
+		skippedDueToFailures := 0
+		for _, visit := range vdb.visitsDB { // visitsDB accessible within package
+			if strings.TrimSpace(visit.Visit) == "" {
+				continue
+			}
+			sentences := extractSentences(visit.Visit)
+			for _, s := range sentences {
+				processed++
+				trimmed := strings.TrimSpace(s)
+				if len(trimmed) < 3 {
+					continue
+				}
+				// Existence check before any LLM calls
+				existsCtx, existsCancel := context.WithTimeout(context.Background(), 2*time.Second)
+				exists, err := repo.ExistsSentenceEmbedding(existsCtx, visit.ID, trimmed)
+				existsCancel()
+				if err != nil {
+					logrus.WithError(err).Warnf("Exists check failed visit=%s sentence=%q", visit.ID, trimmed)
+				} else if exists {
+					skippedExisting++
+					continue
+				}
+
+				// Translation with retry/backoff
+				var translated string
+				translateErr := retry(maxTranslateAttempts, 0, func(at int) error {
+					ctx, cancel := context.WithTimeout(context.Background(), translateTimeout)
+					defer cancel()
+					resp, err := llm.TranslateToEnglish(ctx, trimmed)
+					if err != nil {
+						logrus.WithError(err).Warnf("Translate attempt=%d failed visit=%s sentence=%q", at+1, visit.ID, trimmed)
+						return err
+					}
+					translated = strings.TrimSpace(resp)
+					return nil
+				})
+				if translateErr != nil || translated == "" {
+					translated = trimmed // fallback keep original language
+				}
+
+				// Embedding with retry/backoff (skip if translation totally failed with deadline each time)
+				var emb []float64
+				embErr := retry(maxEmbeddingAttempts, 0, func(at int) error {
+					ctx, cancel := context.WithTimeout(context.Background(), embeddingTimeout)
+					defer cancel()
+					vec, err := llm.GetEmbeddings(ctx, translated)
+					if err != nil {
+						logrus.WithError(err).Warnf("Embeddings attempt=%d failed visit=%s sentence=%q", at+1, visit.ID, trimmed)
+						return err
+					}
+					emb = vec
+					return nil
+				})
+				if embErr != nil {
+					skippedDueToFailures++
+					continue
+				}
+
+				persistCtx, pcancel := context.WithTimeout(context.Background(), 5*time.Second)
+				if err := repo.InsertSentenceEmbedding(persistCtx, visit.ID, trimmed, translated, emb); err == nil {
+					inserted++
+				}
+				pcancel()
+				// Throttle (configurable?)
+				time.Sleep(50 * time.Millisecond)
+			}
+		}
+		logrus.Infof("Sentence embedding backfill complete processed=%d inserted=%d skipped_existing=%d skipped_failures=%d elapsed=%s", processed, inserted, skippedExisting, skippedDueToFailures, time.Since(start))
+	}()
+}
+
+// retry executes fn up to attempts times with exponential backoff starting at base (or 200ms if base==0)
+func retry(attempts int, base time.Duration, fn func(attempt int) error) error {
+	if attempts <= 0 {
+		return nil
+	}
+	if base <= 0 {
+		base = 200 * time.Millisecond
+	}
+	var err error
+	for a := 0; a < attempts; a++ {
+		err = fn(a)
+		if err == nil {
+			return nil
+		}
+		// backoff except after last attempt
+		if a < attempts-1 {
+			backoff := base << a // exponential
+			time.Sleep(backoff)
+		}
+	}
+	return err
+}
+
+// extractSentences splits a block of text into sentence-like units.
+func extractSentences(text string) []string {
+	// First replace newlines with space to keep regex simpler, keep periods.
+	normalized := strings.ReplaceAll(text, "\n", " ")
+	matches := sentenceSplitRegex.FindAllString(normalized, -1)
+	var out []string
+	for _, m := range matches {
+		m = strings.TrimSpace(m)
+		if m != "" {
+			out = append(out, m)
+		}
+	}
+	return out
+}