NutriBot — AI Nutrition Assistant

NutriBot is HealtyHelp’s built-in conversational assistant. It uses the Groq SDK with the LLaMA 3.3 70B Versatile model to answer nutrition and recipe questions, always staying within the context of the user’s personal health profile and the platform’s recipe catalog.

How NutriBot Works

Every time a user sends a message, the server:

Fetches the user’s name, healthProfile, and alergia fields.
Fetches the full recipe catalog, selecting only nombre, desc, cat, salud, ingredientes, and core nutri fields (cal, prot, carb, gras).
Filters recipes by the user’s active categorias (if any) and then removes any recipe whose ingredients contain an allergen from the combined allergy list.
Assembles a system prompt that includes the configurable base prompt, the filtered recipe list, and the user’s full health profile.
Appends the last 10 conversation turns as chat history (user and assistant roles).
Sends everything to Groq and streams back the response.

AI Model Configuration

Parameter	Value
SDK	`groq-sdk` (Node.js)
Model	`llama-3.3-70b-versatile`
`max_tokens`	`1024`
System prompt	Loaded from `AIConfig` MongoDB document (cached for 60 seconds)

Allergy Filtering

Before the recipe list is sent to the AI, any recipe that contains an allergen in its ingredientes array is removed:

function filtrarPorAlergias(recetas, alergias) {
  if (!alergias?.length) return recetas;
  const alergiasNorm = alergias.map(a => a.toLowerCase().trim());
  return recetas.filter(receta => {
    const ingredientes = receta.ingredientes ?? [];
    return !ingredientes.some(ing =>
      alergiasNorm.some(al => ing.toLowerCase().includes(al))
    );
  });
}

The allergy list is the union of two sources on the User document:

healthProfile.alergias[] — set via the health profile form
alergia — the legacy single-value field

NutriBot will never recommend a recipe whose ingredients match any of the user’s allergens, because those recipes are removed before the prompt is built.

System Prompt Structure

The system prompt sent to Groq is assembled at request time and includes:

The configurable base prompt — loaded from the AIConfig collection (see Configuring the AI Prompt).
The filtered recipe list — one line per recipe in the format • Nombre [cat] | Nkcal | desc.
The user’s health profile — conditions, active food categories, allergies, and preferences.
Behavioral rules — constraining the model to only recommend listed recipes and respect medical conditions.

Example System Prompt Excerpt

This is the default prompt stored in the AIConfig document:

Eres NutriBot, el asistente nutricional de HealtyHelp.
Solo respondes preguntas sobre nutrición, recetas y alimentación saludable.
Si el usuario pregunta algo fuera de estos temas, responde amablemente
que solo puedes ayudar con temas nutricionales.
Responde en español, de forma clara, amigable y concisa.

RECETAS DISPONIBLES (42 — filtradas por: desayuno):
• Avena con frutas [desayuno] | 320kcal | Nutritivo desayuno ...
• Tostadas integrales con aguacate [desayuno] | 280kcal | ...
...

Perfil de "María":
- Condiciones / dieta: diabetes, hipertension
- Tipo de comida activo: desayuno
- Alergias: maní
- Preferencias: sin gluten

REGLAS:
- Recomienda SOLO recetas de la lista anterior. Esa lista ya fue filtrada por alergias del usuario.
- Nunca inventes recetas que no estén en esa lista.
- Respeta condiciones médicas.
- Si el usuario pregunta qué filtros tiene activos, responde exactamente con los valores del perfil de arriba.

The last section of the prompt also instructs the model whether this is the first message of a session (so it greets the user by name once) or a continuation (skip the greeting).

Conversation History

The last 10 messages from the client-side history array are included in each request. Messages with role: "model" are mapped to role: "assistant" for the Groq API. The history is trimmed to always start with a user turn:

const chatHistory = history.slice(-10).map(msg => ({
  role: msg.role === 'model' ? 'assistant' : 'user',
  content: msg.text,
}));
// Ensure the history starts with a user message
while (chatHistory.length > 0 && chatHistory[0].role !== 'user') {
  chatHistory.shift();
}

The Two Chat Interfaces

A small robot icon button fixed to the bottom-right corner of every page. Clicking it toggles a compact ChatCore panel. From this panel the user can expand to the full chatbot page via an “Expandir” button that navigates to /chatbot.

// RobotIA.jsx
<button className="robotBoton" onClick={toggleIA} title="Asistente IA">
  <IcoRobotBtn />
</button>
{activo && (
  <div className="robotChat">
    <ChatCore
      modoExpandido={false}
      onExpandir={irAChatCompleto}
      onCerrar={toggleIA}
      ...
    />
  </div>
)}

Full Chatbot Page (`VistaChatbot`)

Accessed at /chatbot. Features a two-column layout:

Left panel — branding, and a list of NutriBot’s topic capabilities (chronic conditions, special diets, adapted recipes, allergies/intolerances, nutritional advice).
Right panel — full-height ChatCore in expanded mode with a “Minimizar” button that returns to the floating widget.

Both interfaces share the same useChatStore hook, so the conversation history is preserved when switching between them.

Health Profile Management

Users configure the profile that NutriBot uses via the health profile settings form, which calls:

PUT /api/chat/health-profile

Request body:

{
  "condiciones":  ["diabetes", "hipertension"],
  "categorias":   ["desayuno", "almuerzo"],
  "alergias":     ["maní", "gluten"],
  "preferencias": ["sin lactosa"]
}

All four fields are arrays of strings. Omitting a field leaves it unchanged. Response:

{
  "message": "Perfil actualizado",
  "healthProfile": {
    "condiciones":  ["diabetes", "hipertension"],
    "categorias":   ["desayuno", "almuerzo"],
    "alergias":     ["maní", "gluten"],
    "preferencias": ["sin lactosa"]
  }
}

To retrieve the current health profile filters:

GET /api/chat/filtros

Returns condiciones, categorias, alergias, and preferencias arrays.

Configuring the AI Prompt

Read the current prompt

GET /api/chat/prompt
Authorization: Bearer <token>

{ "prompt": "Eres NutriBot, el asistente nutricional de HealtyHelp..." }

Update the prompt (authenticated users)

PUT /api/chat/prompt
Authorization: Bearer <token>
Content-Type: application/json

{ "prompt": "Nuevo prompt del sistema..." }

The prompt is stored in the AIConfig MongoDB collection (single document) and cached in memory for 60 seconds (CONFIG_TTL_MS = 60 * 1000). Updating the prompt via the API clears the cache immediately (_configCache = null).

`AIConfig` Model

// server/models/AIConfig.js
{
  prompt: {
    type: String,
    default: `Eres NutriBot, el asistente nutricional de HealtyHelp.
Solo respondes preguntas sobre nutrición, recetas y alimentación saludable.
Si el usuario pregunta algo fuera de estos temas, responde amablemente
que solo puedes ayudar con temas nutricionales.
Responde en español, de forma clara, amigable y concisa.`
  },
  timestamps: true
}

Rate Limit Handling

Groq returns HTTP 429 when the API quota is exceeded. The server catches this (and 503, 500, 400, and quota-related error messages) and returns a user-friendly response instead of an error:

{
  "error": "rate_limit",
  "reply": "Estoy recibiendo muchas consultas ahora mismo. Intenta de nuevo en unos segundos."
}

The client should detect error === "rate_limit" and display the reply text in the chat bubble rather than throwing an error.

What NutriBot Will and Won’t Answer

NutriBot will help with:

Questions about any recipe in the catalog (ingredients, nutrition, preparation)
Recommendations for recipes that fit the user’s health profile
General nutrition guidance (macros, vitamins, minerals, hydration)
Advice tailored to health conditions (diabetes, hypertension, keto, etc.)
Clarifying the user’s active filters and conditions

NutriBot will not:

Answer questions outside nutrition, recipes, and healthy eating
Invent recipes that are not in the filtered catalog
Ignore the user’s allergens (allergen-containing recipes are removed before the prompt)
Provide specific medical diagnoses or replace professional healthcare advice

If asked about off-topic subjects, NutriBot responds politely that it can only assist with nutritional topics.

Getting Started

Core Features

Authentication

Admin Panel

Deployment

NutriBot: AI Nutrition Assistant Powered by Groq

NutriBot — AI Nutrition Assistant

How NutriBot Works

AI Model Configuration

Allergy Filtering

System Prompt Structure

Example System Prompt Excerpt

Conversation History

The Two Chat Interfaces

Floating Widget (`RobotIA`)

Full Chatbot Page (`VistaChatbot`)

Health Profile Management

Configuring the AI Prompt

Read the current prompt

Update the prompt (authenticated users)

`AIConfig` Model

Rate Limit Handling

What NutriBot Will and Won’t Answer

Build docs developers (and LLMs) love

Getting Started

Core Features

Authentication

Admin Panel

Deployment

Documentation Index

​NutriBot — AI Nutrition Assistant

​How NutriBot Works

​AI Model Configuration

​Allergy Filtering

​System Prompt Structure

​Example System Prompt Excerpt

​Conversation History

​The Two Chat Interfaces

​Floating Widget (RobotIA)

​Full Chatbot Page (VistaChatbot)

​Health Profile Management

​Configuring the AI Prompt

​Read the current prompt

​Update the prompt (authenticated users)

​AIConfig Model

​Rate Limit Handling

​What NutriBot Will and Won’t Answer

Build docs developers (and LLMs) love

NutriBot — AI Nutrition Assistant

How NutriBot Works

AI Model Configuration

Allergy Filtering

System Prompt Structure

Example System Prompt Excerpt

Conversation History

The Two Chat Interfaces

Floating Widget (`RobotIA`)

Full Chatbot Page (`VistaChatbot`)

Health Profile Management

Configuring the AI Prompt

Read the current prompt

Update the prompt (authenticated users)

`AIConfig` Model

Rate Limit Handling

What NutriBot Will and Won’t Answer