Framer AI AgentsDesign and publish professional sites with AI
Consilience
A personal Palantir that’s as easy to use as Obsidian
31 followers
A personal Palantir that’s as easy to use as Obsidian
31 followers
Build Your Real Second Brain (Ontology + Obsidian + Agent) A personal Palantir that’s as easy to use as Obsidian. Write Markdown documents or import existing content to build a knowledge graph, then work with an AI agent powered by integrated GraphRAG
How does the knowledge graph stay accurate when your local files change a lot, and is there a way to manually correct nodes or connections the agent gets wrong?
@erturulsivv0dv
Great question. In Consilience, your markdown documents are the interface between you and the AI. You just write notes like you normally would, or upload and convert files, and every time a markdown file changes, Consilience rebuilds the graph entities it needs from that file.
The important part is that your files are the source of truth and the graph is derived from them, not the other way around. So when a note changes, only that note re-extracts, and its old entities and relationships get swapped for the new ones in one clean step, with no stale leftovers. We hash the file content, so files you didn't actually change don't burn any model calls, and renames or deletes update the graph on their own. If you edit a file in another editor, it gets flagged as stale in the file tree, and you re-sync with one click, so you stay in control of when it runs.
On top of that, none of this is a one-time thing. It runs as a knowledge graph lifecycle. Consilience handles work like re-extraction, synonym resolution, relationship suggestions, and topic map (MOC) generation for you, either automatically or by suggesting it first depending on the situation. So even as your documents pile up and change often, the graph doesn't get left behind and keeps itself tidy on its own.
As for fixing what the agent gets wrong, since files are canonical, the simplest way is to just edit or delete the text in the note, and the wrong node or connection goes away with it. You can even ask the agent to do that for you. Beyond that, you can correct things at the entity level right from the graph: retype an entity, merge duplicates, split a bad merge, or reject a suggested connection. Those manual corrections stick, so a later re-extraction won't quietly undo them.
Dropped a messy mix of markdown notes and some scripts into it, and the graph view actually surfaced connections I hadn't noticed before. The local-first angle is a nice touch for anyone wary of cloud uploads.
@ayazkabasadqhc
Love hearing this, thanks for actually putting it through its paces. That moment where the graph points out a connection you didn't know was there is the whole reason we built it, so I'm really glad it clicked for you.
A messy pile is honestly the ideal input. Consilience indexes your scripts into the graph alongside your notes, so code doesn't just sit in a folder as dead weight, it becomes part of the same picture your notes are in. The more mixed and unstructured the stuff you throw at it, the more of those hidden links tend to come out.
And local-first is a core principle for us, not a marketing checkbox. Your files and the graph stay on your machine as plain files you own, with no cloud vault and no forced sync. Really appreciate you giving it a try and taking the time to write this up.
Been waiting for something like this to wrangle my scattered notes and code. The knowledge graph view actually shows connections I would have missed, and running it locally is a huge plus.
@halitm7tv
This is exactly the itch we were trying to scratch, so it means a lot that it landed for you. Scattered notes and code is such a common mess, and almost nothing treats the two as one connected thing, so I'm glad it's already helping you wrangle it.
The connections you would have missed are really the point, and the visual graph is just the surface. Underneath, the agent runs graph RAG over that same structure and uses SPARQL for multi-hop reasoning across your notes and code. So instead of tracking down the relevant pieces and pasting them in yourself, you can just tell the agent what you want and let it walk the graph to assemble the context and surface insights you would not reach by search alone. The more you add, the more of those multi-hop paths open up.
And it's great that running it locally lands as a plus for you, that was a deliberate choice from the start. Today the AI extraction still runs through a cloud model, but we're also considering bring-your-own-key (BYOK) and local LLM support, so the whole pipeline can eventually stay 100% on your machine with nothing leaving it. Thanks a lot for giving it a try and for the kind words.
the local-first framing is what got my attention, but you mentioned extraction still runs through a cloud model today. for something meant to hold your private notes and code, that's the part I'd want spelled out clearly: is it sending raw file content to that cloud model for extraction, or some stripped-down representation. if it's raw content, local-first currently describes where the graph lives, not where your actual notes travel
The graph view actually feels useful here, not just flashy, and watching the agent pull from linked notes when I asked a question was a nice surprise. Wish it handled larger vaults a bit faster though.
Been playing with the graph view for the past hour and the way it pulls connections from imported notes without me tagging anything is genuinely useful, beats my old PKM setup hands down.
The graph view actually feels useful right out of the box, not just a pretty diagram. Having the agent pull answers from my own notes through GraphRAG is the part that sold me on it.