[brainless]

I tried a graph based approach in my previous product (1). I am on a new product now and I came back to SQLite. Initially it was because I just wanted a simple DB to enable creating cross-platform desktop apps.

I realized LLMs are really good at using sqlite3 and SQL statements. So in my current product (2) I am planning to keep all project data in SQLite. I am creating a self-hosted AI coding platform and I debated where to keep project state for LLMs. I thought of JSON/NDJSON files (3) but I am gravitating toward SQLite and figuring out the models at the moment (4).

  1. Previous product with a graph data approach https://github.com/pixlie/PixlieAI
  2. Current product with SQLite for its own and other projects data: https://github.com/brainless/nocodo
  3. Github issue on JSON/NDJSON based data for project state for LLMs: https://github.com/brainless/nocodo/issues/114
  4. Github issue on expanding the SQLite approach: https://github.com/brainless/nocodo/issues/141

Still work in progress, but I am heading toward SQLite for LLM state.

[eyeris]

What sort of issues did you run into with a graph based approach?

[brainless]

My implementation was custom, on top of RocksDB. I found it hard to ask LLM to traverse it. While understanding schema of SQLite or making queries to find information is very easy for LLMs. In most cases schema does not have to be inferred since it is going to be available and this makes the job easier. The graph approach may work well for many use-cases but if we want to store structured information for LLMs then SQLite is really good.