ThoughtDAG: Turning LLM Conversations Into an Editable Graph You Can Prune
Most chat interfaces treat context as a linear scroll you can't touch. ThoughtDAG rebuilds that context as a graph, letting you see, wire, and delete the branches that feed into a model request before you send it.
Most chat interfaces treat context as a linear scroll you can't touch. ThoughtDAG rebuilds that context as a graph, letting you see, wire, and delete the branches that feed into a model request before you send it.
Specs
- License: MIT
- Model integrations: Ollama and OpenAI-compatible endpoints
- Document support: PDF clipping with page provenance
- Context control: Graph wiring and edge deletion
What's new
- Conversation context is represented as an editable graph rather than a linear thread.
- The same prompt can produce different answers depending on how the graph context is wired.
- Deleting an edge removes its branch from the next model request, producing a cleaner answer sourced from 2 sources in the example given.
- Before sending a request, the interface previews which source nodes are included, their ordering, and the total token count.
- PDF passages can be selected and questioned directly, and they retain page provenance within the graph.
- An example research conversation in the project runs to 87 messages, illustrating how large these graphs can grow.
Who is it for
- Researchers and analysts who work through long, branching LLM conversations and need to isolate which context actually shaped an answer.
- People doing document-heavy research who want to pull questions from specific PDF passages while keeping track of page provenance.
- Developers who want a local-first setup that can run against Ollama or any OpenAI-compatible endpoint instead of a single locked-in provider.
- Anyone who has hit the problem of a chatbot answer being quietly distorted by irrelevant earlier context and wants a way to inspect and prune that context directly.
How to try it
ThoughtDAG is an MIT-licensed, local-first application available on GitHub (chenxiachan/thoughtdag). It's built to connect to Ollama or any OpenAI-compatible endpoint, but exact setup steps, backend configuration, and packaging details (including the AppImage build) should be verified directly in the repository's documentation before running it, since community reports have flagged open questions around sandboxing and network binding that aren't yet resolved in the public docs.
Community take
- Several people liked the core idea, comparing it to Git for context history or to manually maintained Markdown decision trees, and specifically praised the semantic zoom, weave, and condense features.
- A number of commenters said they'd rather see this as a plugin or integration inside existing editors, planning tools, and coding assistants like Claude Code, Codex, Pi, and Replit, rather than as a standalone app, and one found the rationale for a standalone app unclear.
- Others offered usability suggestions, including surfacing weave and condense ahead of edge deletion for less technical users, and adding a node sidebar that shows references and lets people remove context directly from there.
Sources
- ThoughtDAG repository — GitHub