Improve LLM-based documentation exploration for Actionbase

[em3s]

Background

Actionbase aims to be easier to approach for both contributors and users.

The Actionbase documentation is publicly available at https://actionbase.io/. As more LLM tools can fetch and read content directly from the web, Actionbase should support workflows that go beyond manual download and upload.

Note

While preparing Actionbase for its open source release, the documentation was revised to be clearer and more consistent. LLM tools were used to review the written content.

The workflow was:

• Bundle documentation as a zip file
• Upload it to an LLM tool
• Ask for review or assistance

This approach works, but it does not scale well. A more flexible and LLM-friendly documentation exploration workflow is needed.

Current Approach

As a first step, Actionbase adopted llms.txt (https://github.com/delucis/starlight-llms-txt) and provides guidance at https://actionbase.io/llms-txt/.

This has established a basic integration point for LLM-based documentation access, but there is still room for improvement.

Goal

The goal of this issue is not to decide on a single direction upfront, but to explore multiple approaches through experimentation and iteration.

Based on these explorations, we would like to evaluate what works best for Actionbase and then discuss whether to:

• improve the existing https://actionbase.io/llms-txt/ page, or
• replace it with a clearer and more structured guide for LLM-based documentation exploration.

Incomplete proposals are welcome, including small experiments or draft-level pull requests.

[em3s]

I tried a few quick experiments to validate this direction.

Test prompt

What is Actionbase, and what problems does it aim to solve?
Answer in 3–5 bullet points.

1. NotebookLM

Setup: Add https://actionbase.io/llms-full.txt as a website source, then run the test prompt.

Result:

Actionbase is a specialized database designed to handle the unique challenges of managing user relationship data at a massive scale. According to the sources, here is what it is and the problems it addresses:
...

2. ChatGPT

I tested the following prompt first:

Read https://actionbase.io/llms-full.txt

You are an assistant answering questions about Actionbase.
Answer using only the provided Actionbase documentation.
If the documentation does not specify the answer, say so explicitly.

Then I applied the test prompt.

Result:

Actionbase is a database for serving user interactions, modeling actions like follows, likes, or reactions as graph edges between entities, exposed through a unified REST API.
...

While the test using llms-full.txt seems to work, across experiments the output tends to focus on the API Reference due to its volume. There may be better ways to approach this. 🤔

[em3s]

To address the API Reference volume issue, I split the documentation:

• core.txt: Full docs without API reference
• api.txt: API reference only

https://actionbase.io/llms-txt/