Building a RAG Knowledge Base with Memories

This is a step-by-step walkthrough: build a memory folder, make it searchable, and query it from a Project and a Task. For the concepts behind each step — what Searchable by AI does, how indexing and propositions work, and who can read a folder — see Memories. This page stays hands-on and links back rather than re-explaining.

Walkthrough

Step 1: Set up the memory folder

1. Go to Memories in the sidebar and create a new folder.
2. Open the folder Settings.
3. Enable Searchable by AI.

That one setting turns the folder into a searchable knowledge base — Aisle indexes every record, including any you added before enabling it. See Making a folder searchable for what happens under the hood.

Step 2: Add your documents

There are two ways to populate the folder:

Method	Details
Upload files	Supported formats: PDF, Word, CSV. Content is extracted automatically.
Create manually	Write markdown content directly in the editor.

Each document is queued for indexing immediately after saving.

Step 3: Wait for indexing

Indexing runs in the background, and each document shows a status — pending, processing, completed, or failed. Only completed documents are returned by vector search; re-save a failed one to retry. See Making a folder searchable for the full status lifecycle.

Step 4: Use the folder in a Project

Assign the folder to a Project as Reference. The model can then search the folder automatically during project chats.

If project work should save new documents into the folder, also assign it as Output.

Ask a question in the project:

Answer this using the Customer Research knowledge base: what renewal risks have we seen for Acme?

Step 5: Search the folder from a Task

Tasks can search the same folder explicitly:

question = aisle.inputs.get("question")

context = aisle.memories.vector_search(
    query=question,
    folder="Customer Research",
    limit=5,
    threshold=0.7,
)

answer = aisle.ai.raw(
    f"Answer using only this context:\n\n{context}\n\nQuestion: {question}"
)

aisle.create_chat("Knowledge base answer", answer)

The query is semantic - searching for "billing problems" will match documents containing "customer charged twice" without requiring keyword overlap.

Step 6: Test it

1. Ask a sample question in the Project, or run the Task with a sample question.
2. Inspect the retrieved context or execution logs to review which propositions were returned and their similarity scores.
3. If results are too broad, increase the similarity threshold. If too few results are returned, lower it.

Going further

• Empty results handling: In a Task, branch when context is empty and return a fallback response.
• Multiple folders: Search separate knowledge bases from the Project or from a Task, depending on how explicit you need the retrieval logic to be.