User repository on GitHub
Typesafe.ai System One Model Jev navigating a Neo4j graph by using a classifier over neighbouring relationships
- Forks
- 27
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Cookbooks
Neo4jev
A demo application that navigates a Neo4j graph one hop at a time: Jev gets the outgoing relationships as a Choice plus a goal Noul in one system_one call, and beam search follows the probabilities.
- Stars
- 174
- Language
- Jupyter Notebook
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Author
- jexp
- Last verified
- Sep 2026
About this project
A demo application that navigates a Neo4j graph one hop at a time using TypeSafe's (typesafe-sdk) systemone structured-decision API instead of free-text LLM generation. At each visited node, the outgoing relationships are presented as Choice options (relationship type, relationship properties, target node label/properties), and the API returns a full probability distribution over which relationship to follow next. A Noul ("has the goal been reached?") question rides in the same systemone call as the Choice, so…
Quoted from the project's README. Read the full README.
Works forJev for Scoring
Used inJev for Software & DevTools
Highlights
- Each hop is one system_one call: the outgoing relationships are Choice options and the goal-reached question is a Noul in the same request.
- Beam search branches on top-k probabilities and ranks chains by summed log-probabilities to avoid float underflow and length bias.
- Three goal modes: free-text goal, explicit target node, and path-intent pattern, which returns the top-N distinct scoring paths.
- Schema-agnostic: labels, relationship types, and property names are discovered by live introspection, so any Neo4j instance can be targeted.
- Results render in an interactive neo4j-viz graph that styles the chosen path separately from the surrounding neighborhood.
Quickstart
bash
uv sync
cp .env.example .env # set NEO4J_* and TYPESAFE_API_KEY
uv run pytest tests/unitWatch out
MIT-licensed. Needs Python 3.12 or newer, uv, a reachable Neo4j instance, and TYPESAFE_API_KEY; without a key each call fails visibly and the pipeline runs on labelled stand-in answers.
Questions about this entry
Short answers drawn from the linked project or page.
No. Each system_one call is attempted, the API's own error is shown verbatim, and the rest of the pipeline runs on explicitly labelled stand-in answers. Nothing is ever presented as TypeSafe output that did not come from TypeSafe. Two consequences are stated in the notebooks: nothing terminates on goal_reached without a key, and target-node mode cannot be demonstrated by a stand-in at all.
It is schema-agnostic. Labels, relationship types, and property names are discovered by live introspection rather than hardcoded, so any instance works through .env variables. It ships against the public Neo4j Companies KG by default; the library also runs headless, with Streamlit only in the app layer.
Each hop is sampled, not exhaustive: at most 10 relationships per type and 60 in total, filled round-robin across relationship types in name order so one large type cannot crowd out the rest. Edges left out can never be chosen, so raise total_cap (a parameter, not a constant) if you need a wider window than the default budget.
Not on the default instance. The server's genai plugin is unconfigured and no provider key is supplied, so neo4j_access.default_embedder is a deterministic hash-seeded placeholder: vector search is dimensionally real but not semantically meaningful. Pass a real embedder to open_access() to change that.
One round-trip per hop. The Choice question over outgoing relationships and the Noul goal-reached question ride in the same system_one call, so more questions do not add round-trips.
Free-text goal, an explicit target node, or a path-intent relationship pattern. Path-intent runs multiple beam branches and returns the top-N distinct scoring paths.
Each hop is capped at 10 relationships per type and 60 in total. The 60-edge budget is filled round-robin across relationship types in name order, so no single type crowds out the rest; raise total_cap to widen the window.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: LinkedIn bookmarks search
X: 1,066 reaction videos, searchable
X: Intent-based search in a Chrome extension
X: Margin, bookmark search
More Jev resources like this
A terminal Wikipedia link-race tool: Chromium fetches article pages, Jev ranks their outgoing links by expected remaining hops, and Python runs a beam search without a search API or a controller model.
Jev Cookbooks#community#python#cli
Community
komikat/jev-bfsPython
komikat/jev-bfsPythonA local 3D reference gallery where Jev chooses search phrases and highlights matching Met, NASA, Cosmos, and Prelinger Archive items from their text metadata.
Jev Cookbooks#community#typescript#demo
Community
AlbionaHoti/refgardenTypeScript
AlbionaHoti/refgardenTypeScriptA community curriculum of Jupyter notebook recipes that teach typed decisions with Jev, TypeSafe's System One model. Each recipe takes one bounded Choice, Noul or Score decision over text or image state, spanning first classification through closed-loop control.
Jev Cookbooks#community#github#python
A small demo app that fetches a URL, turns the page into clean Markdown, and asks Jev ten bounded Choice questions about what a first-time visitor is likely to understand.
Jev Cookbooks#community#typescript#demo
Runnable samples for Jev, TypeSafe AI's System One model: each sample sends state plus questions that carry their own answer options and then branches on the typed value returned, including a README checker and an AGENTS.md readiness scorer.
Jev Cookbooks#community#github#samples
Community
aarora79/jev-samples
aarora79/jev-samplesRunnable Python examples for Jev, TypeSafe's first System One model: the README explains the three typed question primitives — Choice, Score and Noul — gives a venv install for Python 3.10+ and is explicit that the model cannot chat or generate text.
Jev Cookbooks#community#github#python
Community
ignitewala/system-onePython
ignitewala/system-onePythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.