Two of the largest star-gainers on the October 1, 2026 AI open-source trend digests share a premise: for agents, retrieval should follow a document's structure rather than embedding similarity. One tracker estimated Graphify-Labs/graphify at about 1,200 new stars for the day and VectifyAI/PageIndex at about 1,100. These figures are third-party estimates, and the same digest showed obviously wrong totals for some other repos. NVIDIA's OpenShell, covered here yesterday, was also near the top.
Graphify (about 123,000 stars, Apache-2.0 and MIT) builds a knowledge graph from a codebase plus its docs, SQL schemas, configs and PDFs, and installs as a /graphify skill for Claude Code, Cursor, Codex, Gemini CLI and more than a dozen other coding assistants. Code is parsed locally with tree-sitter across 37 language grammars, with "no LLM calls" and nothing leaving the machine. Every edge is tagged EXTRACTED (stated in the source) or INFERRED (resolved by the tool), so an agent can tell which relationships it read and which were guessed. An MCP server exposes tools such as get_neighbors, shortest_path and get_pr_impact.
PageIndex (about 38,000 stars, MIT) is "vectorless" RAG for long professional documents such as filings, contracts and manuals. It builds a tree index from each document's layout, then has an LLM reason its way down the tree to the relevant section, in place of chunking and embedding. Its motto is "similarity ≠ relevance." The project reports 98.7% accuracy on FinanceBench against roughly 50% for vector RAG. That is the vendor's own figure. It ships an MCP server and added a local SDK mode in August.
The common thread: coding and document agents keep failing on retrieval that is semantically close but structurally wrong, such as the function with a similar name or the paragraph from the wrong section of a 10-K. Both projects trade embedding speed for explicit structure that can be traced. The costs are an LLM call per navigation step for PageIndex, and a graph that goes stale as code changes for Graphify. Teams should benchmark them on their own corpora rather than rely on star counts or vendor accuracy figures.
Graphify and PageIndex, two of October 1's biggest gainers, replace embedding similarity with explicit structure -- a provenance-tagged code graph and a reasoning-navigated document tree -- which makes retrieval more traceable, at the cost of extra LLM calls or graph upkeep.