Feed AI

Feed AI is a powerful tool within SmartFeeds that allows you to use artificial intelligence to make many different kinds of changes to your product data. By creating your own "Tasks" and "Cases," you can give the AI specific instructions to create new attributes, rewrite titles and descriptions, or even check the quality of your data.

This guide will show you the process of setting up your first AI Task using a practical example, from selecting your input data to launching the optimization and interpreting the results.

Prerequisites

Before you begin, you should have an active product feed imported into SmartFeeds. You should also be familiar with the general interface of the Product Flow. This is the central area where you manage your data. Understanding features like joining (to combine different data sources) and mapping (to standardize attributes) is key to getting the most out of SmartFeeds.

Looking for inspiration? Click here to access our exclusive library of pre-built prompts and schemas to kickstart your optimizations.

Understanding the Feed AI Interface: Tasks and Cases

The core of Feed AI revolves around two main concepts. Think of it like organizing a project:

A Task is like the main project folder. It holds everything for a specific optimization goal, including the input data and general settings (like the AI model). You can create different Tasks for different projects
Example: one Task (or “folder”) for attribute extraction, another for title optimization
A Case is like a specific instruction sheet inside your project folder. Each Task can contain one or more Cases. A Case is made of three parts that work together: a product filter, a prompt, and a schema. This allows you to apply different instructions to different segments of your products, all within the same project. Cases are processed in order; a product handled by the first instruction sheet will be excluded from the next.
Example : within a single "Title Optimization" Task, you could have one Case with a prompt to create long, detailed titles for your "Electronics" category, and a second Case with a different prompt to create shorter, punchier titles for your "Sport" category.

From the main interface, you can also manage your Tasks to keep your workspace organized:

Duplicate/Delete: You can duplicate or delete an entire Task by clicking the three-dot menu next to the Task name.
Collapse Cases: To keep your workspace tidy when working with multiple Tasks, you can collapse the Cases of any inactive Task.

Configuring Your First AI Task: A Step-by-Step Example

To illustrate the process from start to finish, we will show you a concrete example: optimizing product titles. Keep in mind that this is just one of many possible applications. The true power of Feed AI lies in its flexibility, allowing you to build any data transformation scenario you can imagine.

Part 1: Configure and Preview Your Task (Free for All Licenses)

This first part covers all the steps needed to set up your instructions and check the AI's work. These features are available to all users.

Step 1: Creating a New Task and General Settings

First, navigate to the Data Transform section in your Product Flow and select Feed AI. This opens the Feed AI interface.

1. Create a Task: Click the "New Task" button to create your project folder.

2. Define General Settings: Once your task is created, you define the data foundation and high-level parameters for it:

Choose Input for Task: Think of this as choosing the documents you put inside your project folder. Click into this field to select the columns from your product data that the AI will use as its source of information. You can select columns from your source imports and any formulas created in the General section. Channel-specific formulas cannot be used.
Exemple: For our title optimization, you would want to select columns like title, description, brand, category, and any other attribute like color or material. This gives the AI the best context to create a rich and complete new title.
Unique Identifier Field: Select the attribute that serves as the unique ID for your products (e.g., id, sku). This is crucial for joining the AI-generated results back to your main product feed later. Think of this field as the unique "name tag" for each product. When Feed AI generates new data (like an optimized title), it uses this ID to know exactly which product the new information belongs to. This ensures that the new data can be correctly merged back into your main product list.
Row limit: Set the maximum number of products to process each time you run the task. This is a safety and speed setting. When testing your instructions, a small limit like 20 gives you fast results to check your work. It also controls costs by preventing you from accidentally processing thousands of products with an untested setup. This limit is different from the product filter you will set up later; it's just a cap on how many products are processed in one go.
Temperature: This setting (from 0 to 1) controls the "creativity" of the AI. A lower value produces more consistent and predictable results, while a higher value leads to more varied and sometimes surprising outputs.
It's important to remember that the AI doesn't "think" like a human. It predicts the next word based on probabilities, like a very advanced auto-complete. A low temperature makes the AI always choose the most likely, "safest" word. A high temperature allows it to sometimes pick less likely words, which can look like creativity.
- Use a low temperature (like 0.1 to 0.3) for tasks that need to be precise and factual, such as extracting specific attributes or reformatting data.
- Use a higher temperature (like 0.7 to 0.9) for creative tasks, like writing a completely new marketing description or brainstorming different title ideas.
- If you're unsure, start low. You can always increase the temperature if the results are too repetitive.
Model: Select the AI model you want to use. We recommend sticking with the default, GPT-4o-mini. It provides the best balance of speed, intelligence, and cost for most common tasks. You can choose other models if you have very complex needs, but for most uses, the default is the best choice.

Step 2: Configuring a Case

Now it's time to write your "instruction sheet" (the Case). Click + New Case to add your first one. A Case has three parts that you need to configure.

1. Select products: This is the "who" of your instruction.

Here you decide which products this instruction should apply to. This is different from the "Row limit" which is just a general safety cap. The filter lets you be very specific.

Example 1 (All products): If you want the instruction to apply to all products (up to your "Row limit"), simply enter TRUE in the filter box. This is the default setting when you create a new Case, making it easy to test your prompt on a general selection of products first.
Example 2 (Specific products): If you only want to work on products with short titles, you can use a filter like character_count(title) < 50.

2. Prompt: This is the "what" of your instruction.

Here you tell the AI exactly what you want it to do. To help you get started, the field is pre-filled with a detailed example prompt. Let's break down what it does:

It tells the AI to act like an expert (You are a leading digital marketing expert...).
It gives a clear goal (...optimize product titles for Google Shopping...).
It provides very specific rules under "Key Guidelines" (like length, what to exclude, etc.)

You'll notice some instructions are in all caps, like THIS IS VERY IMPORTANT. This is a technique to tell the AI that this specific rule is a top priority and must be followed.

How to adapt the pre-filled prompt? You don't need to write a new prompt from scratch! The pre-filled text is a great starting point. For our goal of creating a simple French title, you would simply edit the existing text:

Language: Change ENGLISH to FRENCH.
Category: Change APPAREL to the correct category for your products.
Title Structure: Modify the structure rule to match the format you want.
Complexity: You can simplify or remove parts of the "Task Plan" if you want a more direct result.
Simply edit these parts directly in the text box to fit your needs.

3. Schema: This is the "how" of your instruction.

It defines how the AI should format its answer using a simple language called YAML. The purpose of YAML is to organize information using key: value pairs and indentation (spacing). This is very important because it forces the AI to give you a structured, predictable answer, instead of a simple block of text.

Within the schema, you can be very precise. For each column (or "field") you want, you can add a description to give specific instructions, or an example to show the AI exactly what you mean.

Like the prompt, the schema field is also pre-filled with an example. Let's look at what the default schema does so you can adapt it:

The default schema asks for two things: improved_title (a single piece of text) and reasoning_steps (a list of text items, like a bulleted list).
Asking for a list of steps is a good technique. It forces the AI to break down its reasoning into clear points, which often improves the quality and helps you check its work.

How to adapt this schema?
The goal is to understand this structure so you can make it your own. For example, you could:

Adapt the columns: The name of your output column should match what you are asking the AI to do. The default is improved_title, but if your goal is simply to create a new title without the idea of "improvement," changing it to new_title makes more sense. Similarly, if your task was to extract attributes, you might rename it to extracted_attributes. This makes your final output file much easier to understand and use.
Change the structure: If you prefer a single paragraph for the explanation instead of a list, you can modify the schema for reasoning_steps so it asks for a single piece of text (string) instead of a list (array). This choice depends entirely on the kind of output you want. The default array is great for detailed, step-by-step explanations, while changing to string is better for a short, simple summary.
Add more detail: You could add a description under new_title that says "The title must be under 120 characters" to reinforce a rule from your prompt directly in the structure.
Explore other data types: Besides string (for text) and array (for lists), you can also use integer (for whole numbers, like 10), float (for numbers with decimals, like 9.99), or boolean (for true/false values). If you need more advanced structures, you don't need to be a developer. A good tip is to search online for "YAML schema examples for [what you want to do]". For example, "YAML schema examples for product specifications".

Understanding this logic allows you to adapt the schema for any task, simple or complex.

Looking for inspiration? Click here to access our exclusive library of pre-built prompts and schemas to kickstart your optimizations.

Step 3: Preview and Refine Your Results

This is your most important step for checking your work. Click the Refresh preview button to see what the AI has produced based on your instructions.

Use this preview to see if the AI understood you correctly. The interface is split into two parts: your input data on the left, and the AI's output on the right. To make it easier to check your work, you have a few tools at your disposal:

Column Selector: Use the search box to show or hide specific columns. This is useful to compare a specific input (like the original title) with a specific output (like the new_title).
Resizable View: You can drag the vertical bar in the middle to give more space to the input or output columns.
Copy Preview: Use the "Copy preview" button to easily share your results with a colleague for a second opinion.

If the results aren't right, you can go back and adjust your Prompt, Schema, or Temperature settings until the output is perfect.

Part 2: Activate and Use Your Results (Full Feature Access)

Once you are satisfied with your preview, the next steps involve launching the task at scale and using the data. These features are part of the full Feed AI functionality.

Ready to go further? The following steps, Launch and Download Results, are included in the full version of Feed AI. To unlock this functionality and apply your optimizations to your entire catalog, please contact our sales team

Step 4: Launching the AI Task

Once you are satisfied with your preview, you are ready to launch the task. The launch button has two main modes:

Launch X new rows: This is the standard option. It intelligently processes only the new products that meet your criteria since the last time you ran the task. This prevents you from reprocessing the same products and saves time and costs.
Reset & Launch: By clicking the dropdown arrow, you can choose this option to restart the optimization from scratch. It will re-process all products that match your filter (up to the row limit), ignoring any previous runs. This is useful if you have changed your prompt or schema and want to apply the new logic to all products.

Once launched, the optimization runs in the background, so you can safely leave the page.

Step 5: Accessing and Using Your Results

Once the run is complete, the Download Results button becomes active. Clicking it will export a CSV file containing your newly generated data. This file will always include the unique identifier you selected, making it easy to use the results.

Following our example, the downloaded file would contain your product IDs, the new optimized titles, and the reasoning for each change. You could then import this file back into SmartFeeds and use a Join to merge this new data with your main product feed, replacing the old titles with the new, AI-generated ones.
To open up the possibilities, think of this downloaded file as a brand new data source. You could use it to extract attributes like "material" or "style", then join them to your feed to create better product categories, build powerful filters for your campaigns, or even create a custom quality score.

Quick Answers to Common Issues

I have a formula that creates an optimized title for my Google Shopping channel. Why can't I select this "google_shopping_title" attribute as an input for Feed AI?

Feed AI can only access columns from your original source imports and formulas created in the "General" section of the Product Flow. It cannot see formulas that are specific to a particular channel (like Google Shopping, Facebook, etc.). To use the logic from your "google_shopping_title" formula, you must first recreate it as a new formula in the "General" section. Once it exists there, it will appear in the input data selection for Feed AI.

I've updated my prompt and changed my input columns. Should I use 'Launch' or 'Reset & Launch'?

You must use Reset & Launch in this case. When you change important parts of your setup like the prompt or the input columns, this option re-processes all products with your new instructions. The standard "Launch" button would skip already-processed products, so your changes would not be applied to them.

My filter is correct, but some products are not being processed by my Case. Why?

Check the order of the Cases within your Task. Products are processed by the first Case whose filter they match. If a product is processed by Case 1, it will be excluded from Case 2 and all subsequent cases, even if it also matches their filters.

The output format in my preview is wrong. What should I check?

If the columns have the wrong names or the data is messy, the issue is likely in your Schema. Go back and check for typos or incorrect indentation. A small error there can break the whole structure.

The AI's response in my preview is not what I expected. How can I improve it?

If the new title is weak or the AI didn't follow your rules, your Prompt needs to be clearer. Try making your instructions more specific. Adding examples to your prompt is a great way to show the AI exactly what you want.

My preview results are too boring or too weird. What should I do?

If all the generated titles are almost identical or completely random, adjust the Temperature setting. Lower it for more predictable and consistent results, or raise it for more variety and creativity.

For further assistance with SmartFeeds, consider reviewing these articles:

For additional help, feel free to reach out via our Contact Us page.