Keep texts and find them by meaning

A Vector Store keeps texts and finds again the closest ones to a question, by their meaning. Something puts texts in it: a step of a workflow, or an agent. Something finds them later: another step, or the agent.

It is not the way to give an agent its knowledge. For what an agent must know (documents, pages, feeds), use a Knowledge Base. See Where each kind of information goes.

• A workflow open in the editor. See Create a workflow. Or an agent, for the second example.

The four stores are under AI > Vector Stores in the Nodes catalog. They are circles: you plug one into the steps and tools that use it.

The database of the pgvector store is yours: an outside connection. It has nothing to do with the database of your Agentivity workspace.

Each store has an optional diamond named Embeddings. Embeddings turn a text into numbers, so that texts can be compared by meaning. With nothing plugged in, the store uses a small local model, which is free. You can plug Local Embeddings or OpenAI Embeddings to choose. For a collection that already holds texts, keep the same model: a model of another size is refused with a message.

A workflow keeps four product sheets, then finds the laptops that fit a question.

1. Drag In-Memory Vector Store onto the canvas. Give it the Memory key products.

2. Drag a Vector Store step. Set Operation to Insert Documents. In Documents, give a list in JSON:

   [
     { "text": "Acme Book 13: ultralight laptop, 0.9 kg, 899 euros.", "metadata": { "kind": "laptop" } },
     { "text": "Acme Pro 16: workstation laptop, 2.1 kg, 1899 euros.", "metadata": { "kind": "laptop" } },
     { "text": "Acme Dock: docking station, 149 euros.", "metadata": { "kind": "accessory" } },
     { "text": "Acme Mouse: wireless mouse, 29 euros.", "metadata": { "kind": "accessory" } }
   ]

3. Drag a second Vector Store step and link it after the first. Set Operation to Get Many, Query to What is the lightest laptop?, Limit to 2, and Metadata filter to { "kind": "laptop" }.

4. Plug the In-Memory Vector Store into the diamond Vector Store of both steps.

5. Run the workflow.

What we saw: the first step gave inserted 4. The second step gave count 2 and documents: the Acme Book 13 first, with a score, then the Acme Pro 16. The accessories were left aside by the filter.

An agent can search a Vector Store itself, with the AI Agent Vector Store Tool.

1. In the workflow of example 1, keep the step Insert Documents, and add an AI Agent step after it, with its model.
2. Drag AI Agent Vector Store Tool onto the canvas, and plug it into the diamond AI Agent Tools of the agent.
3. Plug the same In-Memory Vector Store into the diamond Vector Store of the tool.
4. Give the tool a Name, for example Products, and a Description that tells the agent what the texts are: “The product sheets of Acme: laptops and accessories.”

The tool gives the agent a search named after it: search_products. The agent uses it when it needs to. The developer ran this with a model: asked “the lightest laptop under 1000 euros”, the agent searched search_products and answered with the Acme Book 13, 0.9 kg, 899 euros. We did not run the agent ourselves.

By default, the tool only searches. To let the agent keep texts itself, turn on Allow storing. The agent then also has a tool named store_ and the name, for example store_customer_notes, and puts in the store what it wants: a note, what it learned.

1. Give the tool the Name Customer notes, and the Description “Notes on our customers.”
2. Turn on Allow storing.
3. Use a store that keeps texts (Qdrant, Pinecone or pgvector) if the notes must last after the execution. With In-Memory, they last one execution.

The developer ran this with a model: the agent stored “Mrs Martin prefers to be called in the morning, before 10”, found it again, and answered “before 10”. We did not run it ourselves.

• In the workflow, the step Insert Documents gives a number in inserted, and the step Get Many gives count and documents. Read them in the panel Outputs or in the Inspector.
• For an agent, the Inspector shows the search that the agent made.

• count is 0. The two steps do not use the same store, or the same Memory key. Plug the same circle into both. Or the Metadata filter keeps nothing: check the spelling of the values.
• The agent never searches. Say in its instructions what the texts are and when to look in them. Check the Description of the tool.
• The texts are gone after the execution. The store is In-Memory. Use Qdrant, Pinecone or pgvector.
• A message about the size of the model. The collection was filled with another embeddings model. Plug the same one, or use another collection.

• Do not use a Vector Store as a Knowledge Base. An agent that must know your policies, guides or pages needs a Knowledge Base: it reads the source and stays up to date. See What a Knowledge Base is for.
• What an agent finds in a store is sent to the AI model, like the rest of the conversation.

Give an agent knowledge and a memory.