Show HN: ThoughtDAG – An editable context graph for LLM conversations
ThoughtDAG, a new "Show HN" project, introduces an editable context graph to visualize and manage the often-hidden conversation history in large language model (LLM) interactions. This tool addresses the common problem of LLMs receiving irrelevant context, offering users granular control over what information feeds into their next prompts. Its novel approach promises more precise, transparent, and reproducible LLM conversations, appealing to anyone frustrated by opaque chat interfaces.
The Lowdown
ThoughtDAG is presented as a solution to a fundamental problem in current LLM chat interfaces: the invisibility and uncontrollability of the conversation context. It argues that while chat history is long, users have no clear view of which parts of that history an LLM actually considers for its next response. ThoughtDAG proposes a visual, graph-based interface where "wires are context," making this hidden memory explicit and editable.
Key aspects of ThoughtDAG's approach include:
- The Problem: Current LLM interfaces show only the conversation transcript, not the active context, leading to "polluted" answers when irrelevant historical elements are inadvertently included.
- Externalization: Users can "ask from the source" by selecting passages or figures from documents, turning them into source-linked nodes, ensuring provenance and allowing precise context wiring.
- Inspection: Before sending a prompt, ThoughtDAG allows users to preview what the model will receive, including source nodes, their order, and the total token count, removing guesswork from context management.
- Editing: The core innovation lies in the ability to edit the context graph by deleting edges (connections). This allows users to remove irrelevant branches from the conversation history, ensuring they do not enter the next request.
- The Protocol: ThoughtDAG positions the graph itself as the context protocol, emphasizing that users dictate what enters and leaves the LLM's consideration, offering visibility, editability, and inspectability.
The project demonstrates how a seemingly minor edit, like removing an "unrelated branch" about dinner plans from a research summary conversation, can significantly alter the LLM's response, reduce token usage, and lead to a cleaner, more relevant answer. By making context explicit and modifiable, ThoughtDAG aims to empower users with unprecedented control over their LLM interactions.