Navigating Neo4j with Jev

One line in a roundup. The post is a list of fourteen community projects, and this one gets a single sentence: Jev gives each candidate edge a probability, then beam search steps toward the most likely path. Schema, query style, latency and result quality all live in the linked repo, not in the post.

One per hop, by construction. Each node turn becomes a Choice over that node's candidate relationships, and the returned probability decides which branches the beam keeps. Spend therefore scales with hops times branch factor rather than graph size, and the post publishes none of those counts for the linked project.

Ranking, not retrieval. It scores edges you have already fetched, so it cannot recover a relevant edge the candidate query never returned - the same recall ceiling that limited keyword search in the measured reranking work. The post does not discuss recall at all, which is the first thing to check.

None are measured here. The post cites other people's numbers without conditions: about 7.1 s for a Google Flights run, roughly 60 percent cost reduction on 237 replayed coding turns, and about 9Hz decision rate for the FPS demo. All are author-reported by those projects' own authors, with no shared harness.

The account 思维怪怪, posting as @0xLogicrw, which writes about AI. He notes the waitlist moved fast for him, saying he applied yesterday and was on Jev today, and he compiled the list for people who have the API but are unsure what to build. The post itself is written in Chinese.

Fourteen, each with a one-line description and a GitHub link. They span browser agents, MCP servers for Claude Code and Codex, a Unix CLI, context management, code-review triage, semantic code search, desktop automation, a quant-research project and a model router, so it is a survey of launch-week shapes rather than a benchmark.

He reduces it to a loop: repeatedly ask yes or no, which one, what score, and what to do next. When a task actually needs generating code, writing prose or complex reasoning, his advice is to hand it to a conventional large model. That is the same decision-versus-generation split the rest of the ecosystem describes.