This Repo Builds a Knowledge Graph From Any Document With Nothing but a 7B Model
This is the most precisely accurate one yet — it reads like someone paraphrased the README's own methodology section almost step-by-step.
Confirmed: the repo is rahulnyk/knowledge_graph, described as "Convert any Corpus of Text into a Graph of Knowledge."
"Chunks your document, asks a 7B model which concepts relate" — confirmed exactly. Step 1 of the documented method: split the corpus of text into chunks, assign a chunk_id to each. Step 2: for every text chunk extract concepts and their semantic relationships using an LLM — specifically Mistral 7B OpenOrca, a 7B model, run locally via Ollama.
"Adds an edge between any two names that shared a chunk, right or wrong" — confirmed, and this is the sharpest part of your summary. This describes a distinct third step, separate from the LLM extraction: the method also assumes concepts occurring in the same text chunk are related purely by contextual proximity — a co-occurrence heuristic with its own weight, independent of whether the LLM validated any actual relationship. That's exactly the "right or wrong" caveat you flagged.
"Repeated pairs collapse into one edge with summed weight and concatenated relations" — confirmed nearly verbatim. The README's step 4 states this almost word for word: group similar pairs, sum their weights, and concatenate their relationships, so there's only one edge between any distinct pair of concepts.
"networkx layers in degree and community detection so you can rank concepts by importance" — confirmed. The project calculates the degree of each node and communities of nodes for sizing and coloring in the visualization, explicitly framed as a way to understand how important a concept is to the body of work.
Pyvis force-directed rendering — confirmed as the visualization library, consistent with the interactive physics-based graph shown in the project's own demo.
License — confirmed MIT.
Star count — off by a noticeable margin. The repo currently shows 3.4k stars, not 3.7k. Not a huge gap, but it's the one number in your post that doesn't hold up against a direct check right now — worth using "3.4k" or just "3k+" if you want it to stay accurate as stars fluctuate.