fix(storefront-chat): ground AI replies in catalog and stop KB noise

- Filter stopwords from retrieval term matching so generic words do not
  score unrelated KB chunks.
- For product-availability questions with no product or KB hits, return
  a deterministic cannot-confirm reply (skip LLM).
- Tighten system prompt: no invented inventory, menus, or URLs; lower
  temperature.

Made-with: Cursor

src/lib/ai/storefrontGroq.ts

@@ -7,6 +7,16 @@ type HistoryMessage = {
     content: string
 }
 
+/** Customer is asking whether the store carries / sells a product (not policy/shipping). */
+function isAskingAboutProductAvailability(query: string) {
+    return /\b(do you have|have any|have you got|carry|carrying|sell|selling|in stock|out of stock|stock|available|got any|do you sell|do you carry|looking for|can i (get|buy)|can you get)\b/i.test(
+        query.trim()
+    )
+}
+
+const NO_EVIDENCE_PRODUCT_REPLY =
+    "I do not see that item in this store’s catalog or FAQs I have access to, so I can’t confirm we carry it. If you have another product name or category from this shop, I can check again."
+
 export async function generateStorefrontReply(params: {
     storeId: string
     query: string
@@ -17,6 +27,34 @@ export async function generateStorefrontReply(params: {
     const { storeId, query, visitorId, history } = params
 
     const retrieval = await getBlendedStorefrontContext({ storeId, query })
+
+    // Hard guard: do not invent inventory. If they ask “do you have X?” and we have no product row and no KB/FAQ hit, answer without the LLM.
+    if (
+        isAskingAboutProductAvailability(query) &&
+        retrieval.productCount === 0 &&
+        retrieval.kbCount === 0
+    ) {
+        logger.info("[STOREFRONT_AI_REPLY]", {
+            storeId,
+            visitorId,
+            queryLength: query.length,
+            kbHits: retrieval.kbCount,
+            productHits: retrieval.productCount,
+            latencyMs: Date.now() - startedAt,
+            fallbackUsed: true,
+            groundedSkipLlm: true,
+            error: null,
+        })
+        return {
+            content: NO_EVIDENCE_PRODUCT_REPLY,
+            meta: {
+                kbHits: retrieval.kbCount,
+                productHits: retrieval.productCount,
+                fallbackUsed: true,
+            },
+        }
+    }
+
     const contextLines = retrieval.items.map(
         (item, idx) => `[${idx + 1}] (${item.source}) ${item.title}: ${item.content}`
     )
@@ -30,12 +68,13 @@ export async function generateStorefrontReply(params: {
     })) as GroqChatMessage[]
 
     const system = [
-        "You are JUNO storefront assistant for an ecommerce store.",
+        "You are the JUNO storefront assistant. You ONLY answer using the Context block below (products, FAQs, knowledge chunks).",
-        "Use the provided context first (products, FAQs, knowledge).",
+        "If the Context does not clearly support an answer, say you cannot find it in the catalog or docs you have—do NOT guess.",
-        "If context is insufficient, ask one concise clarifying question.",
+        "Never claim a product exists, is in stock, or that the store carries a category unless a Context line explicitly supports it.",
-        "Do not invent exact prices, stock or policy details that are not present.",
+        "Never invent website layout: no tabs, menus, dropdowns, or navigation paths unless they appear verbatim in Context.",
-        "Keep responses concise, friendly, and sales-supportive.",
+        "Never invent URLs or deep links unless they appear in Context.",
-        "Return plain text only.",
+        "Do not affirm availability with “yes” unless a PRODUCT line or FAQ/KB line in Context clearly covers what the customer asked.",
+        "Be concise and friendly. Plain text only.",
     ].join(" ")
 
     const user = [
@@ -53,13 +92,13 @@ export async function generateStorefrontReply(params: {
     ]
 
     const out = await groqChatCompletion(messages, {
-        temperature: 0.2,
+        temperature: 0.1,
         maxTokens: 420,
         timeoutMs: 28_000,
     })
 
     const fallback =
-        "Thanks for your message. I do not have enough details right now. Could you share the product name or what exactly you want to know?"
+        "I don’t have enough in the store’s catalog or FAQ context to answer that accurately. What product name or topic from this shop should I look for?"
 
     const content = (out.content || "").trim() || fallback
 

src/lib/storefront-chat/retrieval.ts

@@ -7,12 +7,128 @@ type RetrievedContextItem = {
     score: number
 }
 
+/** Strip generic words so "do you have socks?" does not match random KB via "you"/"do". */
+const STOPWORDS = new Set([
+    "the",
+    "a",
+    "an",
+    "and",
+    "or",
+    "but",
+    "in",
+    "on",
+    "at",
+    "to",
+    "for",
+    "of",
+    "as",
+    "by",
+    "with",
+    "from",
+    "is",
+    "are",
+    "was",
+    "were",
+    "be",
+    "been",
+    "being",
+    "have",
+    "has",
+    "had",
+    "do",
+    "does",
+    "did",
+    "will",
+    "would",
+    "could",
+    "should",
+    "may",
+    "might",
+    "must",
+    "can",
+    "this",
+    "that",
+    "these",
+    "those",
+    "i",
+    "me",
+    "my",
+    "you",
+    "your",
+    "it",
+    "its",
+    "we",
+    "our",
+    "they",
+    "them",
+    "their",
+    "what",
+    "which",
+    "who",
+    "whom",
+    "whose",
+    "where",
+    "when",
+    "why",
+    "how",
+    "all",
+    "each",
+    "every",
+    "both",
+    "few",
+    "more",
+    "most",
+    "other",
+    "some",
+    "such",
+    "no",
+    "nor",
+    "not",
+    "only",
+    "own",
+    "same",
+    "so",
+    "than",
+    "too",
+    "very",
+    "just",
+    "any",
+    "there",
+    "here",
+    "about",
+    "into",
+    "through",
+    "during",
+    "before",
+    "after",
+    "above",
+    "below",
+    "between",
+    "under",
+    "again",
+    "further",
+    "then",
+    "once",
+    "please",
+    "thanks",
+    "thank",
+    "hello",
+    "hi",
+    "hey",
+    "yes",
+    "no",
+    "tell",
+    "give",
+    "get",
+    "got",
+])
+
 function tokenize(text: string) {
     return text
         .toLowerCase()
         .split(/[^a-z0-9]+/g)
         .map((t) => t.trim())
-        .filter((t) => t.length >= 2)
+        .filter((t) => t.length >= 2 && !STOPWORDS.has(t))
 }
 
 function scoreByTerms(text: string, terms: string[]) {