Looking to exchange ideas : vault structure layer for LLM context injection and vault maintenance

[gauthierae]

Hi Romain — I came across RTFM via your Obsidian Forum post about navigating a 1,700-note research vault.

I'm working on something adjacent: a local-first middleware that exposes vault structure (not content retrieval) to LLMs via MCP — specifically backlink graphs, orphan detection, health scoring, and context assembly. The idea is the LLM knows which notes are connected before deciding what to retrieve to build context. I also have ideas for keeping the vault's content healthy.

RTFM handles retrieval well. PKMS would complement it as context injection and vault maintenance tools. They might complement interestingly.

I will definitely install RTFM and try it out. I will definitely share feedback.

I am still in the early stages of my project and I am setting up a short online questionnaire. I'll share the link when I am ready if you don't mind.

P.S. You can check my very early prototype (more a POC at this point): https://github.com/gauthierae/PKMS

[roomi-fields]

Hi Gauthier,

Thanks a lot for reaching out — and for taking the time to read through the Obsidian Forum thread before pinging. That kind of careful approach is rare these days.

Honest take, peer to peer:

Where it's the same territory
Two of your three current tools (search_notes, get_note) overlap with what RTFM does today — FTS5 + optional embeddings, plus an expand that returns full content. So on the retrieval surface, we're on the same ground rather than complementary.

Where it gets interesting

• get_backlinks is exactly where I want to push RTFM next. Plan is a generic edges(source, target, relation_type) table in SQLite, populated by each parser at sync time — Markdown wikilinks, Python imports, LaTeX \cite{}, etc. Same SQLite, same MCP, no Neo4j. We'd benefit from comparing approaches on the Markdown side; you may already have working wikilink extraction code.
• The part of your roadmap that is genuinely complementary to RTFM is vault maintenance / hygiene — orphans, health scoring, dead links, duplicate detection. RTFM is purely a retrieval layer for agents and has no ambition there. If PKMS leans into that angle (diagnostics for the human curator, not just retrieval for the LLM), it carves out real, distinct value.

On my side
I have a few small Obsidian-side tools you might find adjacent:

• https://github.com/roomi-fields/in-note-link-suggestions — suggests where to insert [[wikilinks]] based on frontmatter (title, focus keyword, tags) or Smart Connections
• https://github.com/roomi-fields/obsidian-content-publisher — publish notes to Substack, WordPress, LinkedIn from Obsidian
• https://github.com/roomi-fields/obsidian-image-autocrop — auto-crop, square and resize PNG images in target folders (handy for AI-generated illustrations)
• https://github.com/roomi-fields/Obsidian-Surfing — fork of the Surfing plugin that adds the ability to run user scripts inside browser views

Collab
Genuinely open to it if a concrete shape emerges — happy to be a beta tester on PKMS once you have something installable, and to share the graph design once I start prototyping it on the RTFM side. Send the questionnaire whenever it's ready, I'll fill it.

Good luck with the project.

— Romain

[gauthierae]

Hi Romain,

Your reply is greatly appreciated.

You are right about the overlap: this was my proof of concept, so I aimed for something relatively simple.

Here is the link to the survey : https://tally.so/r/7RJ1EA.

I will also take a look at your other Obsidian-side tools, it looks interesting.

Hopefully in the coming week or two I can get something installable.
I will definitely let you know.

Alain