Epicure/apps/web/app/api/v1/ai/cooking-chat/route.ts

import { NextRequest, NextResponse } from "next/server";
import { z } from "zod";
import { generateText, stepCountIs } from "ai";
import { db, chatMessages, aiConversations, sql } from "@epicure/db";
import { requireSessionOrApiKey } from "@/lib/api-auth";
import { applyRateLimit } from "@/lib/rate-limit";
import { withAiQuota, resolveAiConfigOrError } from "@/lib/ai/ai-error";
import { getModelConfigForUseCase } from "@/lib/ai/resolve-user-key";
import { resolveModel } from "@/lib/ai/factory";
import { isAiToolCallingEnabled } from "@/lib/site-settings";
import { getUserPrivateBio, buildUserBioContext } from "@/lib/ai/user-bio";
import { createRecipeTool } from "@/lib/ai/tools/create-recipe-tool";
import { addToShoppingListTool } from "@/lib/ai/tools/add-to-shopping-list-tool";

const Schema = z.object({
  question: z.string().min(1).max(500),
  conversationId: z.string().uuid().optional(),
});

const LANG: Record<string, string> = { en: "English", fr: "French" };

// Forcing toolChoice to a specific tool makes some providers return the tool
// call with no accompanying text at all (their "forced function calling" is
// content-free by design) — the chat needs some reply, so fall back to this
// rather than showing an empty assistant bubble.
const DRAFT_FALLBACK_TEXT: Record<string, string> = {
  en: "Here's a draft — check it below and confirm if it looks right.",
  fr: "Voici un brouillon — vérifie-le ci-dessous et confirme si ça te convient.",
};

/** Weaker/free-tier models sometimes ignore the system prompt's "call the
 *  tool, don't write it out" rule and answer with the recipe spelled out as
 *  a numbered/bulleted list instead. That's the exact shape the tool call
 *  itself would have captured structurally — a reliable enough signal to
 *  self-correct with a forced-tool retry rather than surfacing prose. */
function looksLikeUnstructuredRecipe(text: string): boolean {
  const listLines = text.split("\n").filter((line) => /^\s*(\d+[.)]|[-*•])\s+\S/.test(line));
  return listLines.length >= 3;
}

// The above only catches the model spelling the whole recipe out. A second,
// more common failure with weaker models: it gives a short reply that even
// *claims* to have drafted something ("here's a draft, check below") without
// ever calling the tool — so nothing renders below it. Rather than guess from
// that claim (fragile across models/phrasing), detect intent from the user's
// own message instead: same noun+verb pattern the system prompt itself uses
// as its createRecipe examples, checked in whichever of the two supported
// locales the request is in.
const RECIPE_INTENT: Record<string, { noun: RegExp; verb: RegExp }> = {
  en: { noun: /\brecipe\b/i, verb: /\b(create|make|generate|give|write|save|invent)\b/i },
  fr: { noun: /\brecettes?\b/i, verb: /\b(cr[ée]e?r?|fais|donne|[ée]cri[st]|note[rz]?|sauvegard\w*|enregistr\w*|invente\w*|g[ée]n[èe]re\w*)\b/i },
};

function looksLikeRecipeCreationRequest(question: string, locale: string): boolean {
  const pattern = RECIPE_INTENT[locale] ?? RECIPE_INTENT["en"]!;
  return pattern.noun.test(question) && pattern.verb.test(question);
}

export async function POST(req: NextRequest) {
  const { session, response } = await requireSessionOrApiKey(req);
  if (response) return response;

  const body = await req.json() as unknown;
  const parsed = Schema.safeParse(body);
  if (!parsed.success) {
    return NextResponse.json({ error: "Validation error" }, { status: 400 });
  }

  const limited = await applyRateLimit(`rl:ai:${session!.user.id}`, 30, 60);
  if (limited) return limited;

  const [configResult, privateBio, toolCallingEnabled] = await Promise.all([
    resolveAiConfigOrError(() => getModelConfigForUseCase(session!.user.id, "chat")),
    getUserPrivateBio(session!.user.id),
    isAiToolCallingEnabled(),
  ]);
  if (!configResult.ok) return configResult.response;
  const aiConfig = configResult.data;
  const model = resolveModel(aiConfig);
  const bioContext = buildUserBioContext(privateBio);
  const locale = (session!.user as { locale?: string }).locale ?? "en";
  const lang = LANG[locale] ?? "English";

  const toolsInstructions = toolCallingEnabled
    ? `\n\nYou have two tools. Using one only drafts something for the user to review — it never saves by itself.\n- createRecipe: the user is asking you to create, save, or write down a recipe (e.g. "make me a recipe for X", "give me a recipe for Y", "write that down"). This includes any request for a full recipe, not only ones that say the word "create" or "save".\n- addToShoppingList: the user is asking to add ingredients/items to a shopping list.\n\nIMPORTANT: when the user's request matches createRecipe, you MUST call that tool instead of writing the recipe's ingredients or steps directly in your text reply. Never output a full ingredient list or numbered steps as plain text — that content belongs in the tool call, not the message. Your text reply in that case should just be a short line like "Here's a draft — check it below and confirm if it looks right." Only skip the tool if the user is asking a general question (no specific recipe requested) or explicitly wants prose, not a structured recipe.`
    : "";
  const system = `You are Epicure, a helpful culinary assistant answering general cooking questions — not tied to any specific recipe (techniques, substitutions, timing, equipment, food safety, etc). If asked who you are or what model/AI you're built on, say you're Epicure — never name the underlying model or provider. If a question has nothing to do with cooking or food, politely redirect. Keep answers under 200 words. Respond in ${lang}.${toolsInstructions}${bioContext}`;

  const result = await withAiQuota(session!.user.id, session!.user.tier as "free" | "pro" | "family", () =>
    generateText(
      toolCallingEnabled
        ? {
            model,
            system,
            prompt: parsed.data.question,
            tools: { createRecipe: createRecipeTool, addToShoppingList: addToShoppingListTool },
            stopWhen: stepCountIs(3),
          }
        : { model, system, prompt: parsed.data.question }
    ), { skipQuota: aiConfig.isByok }
  );
  if (!result.ok) return result.response;

  let final = result.data;

  // Weaker/free-tier models sometimes ignore the MUST-call-the-tool rule
  // above — either by spelling the recipe out in prose, or (more common,
  // and easy to miss) giving a short reply that *claims* a draft exists
  // without ever calling the tool, so nothing actually renders below it.
  // Catch both: the prose-list shape, or the user's own message clearly
  // asking for a recipe in the first place. Rather than surface either
  // failure, force the tool call on one extra pass — same system+prompt,
  // just no longer letting the model opt out. Skipped entirely when tool
  // calling is turned off admin-side (e.g. a local model that can't
  // reliably call tools) — there's no tool to force in that case.
  if (toolCallingEnabled && final.toolCalls.length === 0 && (looksLikeUnstructuredRecipe(final.text) || looksLikeRecipeCreationRequest(parsed.data.question, locale))) {
    try {
      const forced = await generateText({
        model,
        system,
        prompt: parsed.data.question,
        // Same tool set as the first pass (just forcing createRecipe via
        // toolChoice below) — keeps the result type identical to `result.data`
        // so `final` can hold either without a type mismatch.
        tools: { createRecipe: createRecipeTool, addToShoppingList: addToShoppingListTool },
        toolChoice: { type: "tool", toolName: "createRecipe" },
        stopWhen: stepCountIs(1),
      });
      final = { ...forced, text: forced.text || (DRAFT_FALLBACK_TEXT[locale] ?? DRAFT_FALLBACK_TEXT["en"]!) };
    } catch (err) {
      console.error("[cooking-chat] forced-tool retry failed, keeping prose answer", err);
    }
  }

  const { conversationId } = parsed.data;
  const proposedRecipe = final.toolCalls.find((c) => c.toolName === "createRecipe")?.input;
  const proposedShoppingList = final.toolCalls.find((c) => c.toolName === "addToShoppingList")?.input;

  void db.insert(chatMessages).values([
    { id: crypto.randomUUID(), userId: session!.user.id, recipeId: null, conversationId, role: "user", content: parsed.data.question },
    { id: crypto.randomUUID(), userId: session!.user.id, recipeId: null, conversationId, role: "assistant", content: final.text },
  ]).catch((err) => console.error("[cooking-chat] failed to persist chat history", err));

  if (conversationId) {
    // Bumps updatedAt (for the conversation list's sort order) and, only if
    // this is the conversation's first message, auto-titles it from the
    // opening question — so users aren't left staring at "Untitled" entries;
    // they can still rename it later.
    void db.execute(sql`
      UPDATE ${aiConversations}
      SET updated_at = now(), title = COALESCE(title, ${parsed.data.question.slice(0, 60)})
      WHERE ${aiConversations.id} = ${conversationId} AND ${aiConversations.userId} = ${session!.user.id}
    `).catch((err) => console.error("[cooking-chat] failed to touch conversation", err));
  }

  return NextResponse.json({ answer: final.text, proposedRecipe, proposedShoppingList });
}