Files
Epicure/apps/web/app/api/v1/ai/cooking-chat/route.ts
T
Arnaud 6ec9b65c24 fix: self-heal chatbot recipe requests answered as prose, not a tool call (v0.57.1)
The system prompt already told the model to MUST call createRecipe
instead of writing the recipe out — that's a soft rule some models
(especially free-tier ones) still ignore. There was no fallback: if the
model answered in prose, that's just what got returned, no draft card.

Added a detector (looksLikeUnstructuredRecipe: 3+ numbered/bulleted
lines with no tool call) and a forced-tool retry when it fires — same
system+prompt, but toolChoice forced to createRecipe so the model can't
opt out on the second pass. Falls back to a short localized line if the
provider returns no text alongside a forced tool call (some do this by
design). Retry failure just keeps the original prose answer rather than
erroring the whole request.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
2026-07-20 19:33:49 +02:00

125 lines
7.2 KiB
TypeScript

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 { 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;
}
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] = await Promise.all([
resolveAiConfigOrError(() => getModelConfigForUseCase(session!.user.id, "text")),
getUserPrivateBio(session!.user.id),
]);
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 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 ${lang}.\n\nYou have two tools. Using one only drafts something for the user to review — it never saves by itself.\n- createRecipe: the user is asking you to create, save, or write down a recipe (e.g. "make me a recipe for X", "give me a recipe for Y", "write that down"). This includes any request for a full recipe, not only ones that say the word "create" or "save".\n- addToShoppingList: the user is asking to add ingredients/items to a shopping list.\n\nIMPORTANT: when the user's request matches createRecipe, you MUST call that tool instead of writing the recipe's ingredients or steps directly in your text reply. Never output a full ingredient list or numbered steps as plain text — that content belongs in the tool call, not the message. Your text reply in that case should just be a short line like "Here's a draft — check it below and confirm if it looks right." Only skip the tool if the user is asking a general question (no specific recipe requested) or explicitly wants prose, not a structured recipe.${bioContext}`;
const result = await withAiQuota(session!.user.id, session!.user.tier as "free" | "pro" | "family", () =>
generateText({
model,
system,
prompt: parsed.data.question,
tools: { createRecipe: createRecipeTool, addToShoppingList: addToShoppingListTool },
stopWhen: stepCountIs(3),
}), { 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 and just write the recipe out in prose instead. Rather than
// surface that (or fail the request), force the tool call on one extra
// pass — same system+prompt, just no longer letting the model opt out.
if (final.toolCalls.length === 0 && looksLikeUnstructuredRecipe(final.text)) {
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 });
}