User repository on GitHub
Jev (TypeSafe) 性能評価プロジェクト — 日本郵便 KEN_ALL をマスタに、AI SDK 経由の Jev が住所のあいまい一致にどこまで使えるかを検証
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Practices & Patterns
Jev Jp Address
CLI that normalizes messy Japanese addresses to postal codes using Japan Post KEN_ALL and JIGYOSYO master data, asking Jev only for the steps rules and lookups cannot settle, and printing the probability behind each decision.
- Stars
- 1
- Language
- TypeScript
- Published by
- Community
- Author
- smasato
- Last verified
- Oct 2026
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- TypeScript
Highlights
- Resolves prefecture, city, town, block and building in order, then outputs the postal code, preferring building-level or large-office numbers.
- Jev only ever sees Choice questions over a candidate list plus a none option, with each candidate's KEN_ALL kana reading passed as the criteria description.
- Two-stage fuzzy choice: a first pass including none, then a second over the top 30 similar candidates, adopted only at p≥0.6 and above the first pass.
- npm run robustness scores rules-only against Jev on 218 seeded typo, transposition and kana variants of 28 addresses.
Watch out
The README is candid about what it cannot resolve: input giving only a town name with no city, numeric-only town names such as a Sapporo address written as digits, and Kyoto addresses naming only a street are returned for review or left unresolved. Large-office postal numbers need the JIGYOSYO file present, and without it they are simply not resolved. The default review threshold is 0.7 confidence, and anything below it is flagged with reasons such as multiple postal-code candidates, a conflict with a postal code in the input, or an unidentified town. Jev is described as an early-access model whose probabilities are rounded to two decimals, which is worth keeping in mind when reading the calibration discussion. Accuracy figures are pointed at in eval/report.md and eval/robustness.md rather than printed, and the only cost figure given is one worked example at 3 calls, 3009 tokens and 440 ms. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
Because Jev does not generate text, so the normalized string is always assembled in code. The pipeline runs prefecture to city to town to block and building, consulting Jev at each stage only when forward matching, postal-code lookup and master-data comparison come up short. What it receives is a Choice question over a candidate list plus a none option, with the original address and the previous stage's result in state, and the KEN_ALL kana reading for each candidate passed as criteria to improve matching on kana input.
There are two mechanisms against that. When leftover text after a forward-matched town prefix is not a block number, as in 宮町西2-24 matching 宮町, the rule candidate plus the 30 most similar candidates go to Jev and a different candidate at p≥0.7 overrides the rule match. And if a first Choice pass including none scores below 0.7, a second pass over the top 30 kanji and reading matches must pick the nearest one, adopted only if it reaches 0.6 and beats the first pass. Otherwise the first result comes back flagged needsReview with reasons.
The README gives exactly one worked example, in which a full kana address resolves with 3 calls, 3009 tokens and 440 ms, and the JSON output carries ai.calls, ai.inputTokens and ai.latencyMs so you can measure your own runs. That is the easy case, though: the two-stage fuzzy retry and the rule-verification pass can each add calls, and the README publishes no average or cost per address.
Four cases are called out: input that gives only a town name with no city name; numeric-only town representations, with 札幌市西区24-2-2-3-3 as the example; Kyoto addresses that name a street but no town, since KEN_ALL has no town for them; and anything whose confidence falls under the threshold, default 0.7. Low confidence, multiple postal-code candidates, a conflict with a postal code in the input, and an unidentified town all populate needsReview and reasons.
It needs the Japan Post KEN_ALL and JIGYOSYO master CSVs, fetched by scripts/download-data.sh, and Node.js 20 or newer. Dependencies are only ai and @ai-sdk/typesafe-ai. Run npm install, set TYPESAFE_AI_API_KEY only if using Jev, then npm run build.
The normalized address, the final postalCode with its kind (building, office or town), townPostalCode, and each stage's resolve method (rule, zip, ai or none). An ai stage also carries probability and confidence. needsReview and reasons flag low confidence, multiple candidates, input contradictions and unresolved town names.
If the first pass with a none option returns p below 0.7, a second pass forces the nearest candidate by character bigram and reading similarity from the top 30, with no none option. The second result is accepted only if p is at least 0.6 and higher than the first, otherwise the first pass is returned for review.
More Jev resources like this
7GitHub stars
Word-level language model experiment that uses Jev as its output layer, with explicit tokenizer, sampler, repetition mask, stop rules, and chunk verification.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
y0usaf/jev-lmTypeScript
y0usaf/jev-lmTypeScript30GitHub stars
Chat client over jev-latest, which cannot generate text: the transcript becomes the state and each step is a choice over a 20,000-word dictionary, so replies are built one word at a time.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
Bewinxed/jevgptTypeScript
Bewinxed/jevgptTypeScript3GitHub stars
Plugin for Claude Code and native Codex that layers Jev-assisted task routing, evidence search, context compaction and readable tool output on top of the host, where typed judgments inform policy but the host keeps ownership of permissions and execution.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
MongLong0214/jev-gateTypeScript
MongLong0214/jev-gateTypeScript0GitHub stars
A small Node.js TypeScript CLI for trying Jev on synthetic support text: one urgency question, three questions in a single request, or a latency comparison. No web server and no UI.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
Menny1337/jev-labTypeScript
Menny1337/jev-labTypeScript32GitHub stars
CLI that calibrates Jev questions against your own labels: write questions, label examples in JSONL, run check, and get a verdict per question — gate, gate-above-confidence, ranker, unusable, or too-few-examples.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
smkrv/jev-calibrateTypeScript
smkrv/jev-calibrateTypeScript24GitHub stars
Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rorshopping/jev-on-a-laptopPython
rorshopping/jev-on-a-laptopPythonFrom the Jev guides
Original write-ups on what this entry is for.
Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.
Jev Practices & Patterns#state#design#primitives
Read article
A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.
Jev Practices & Patterns#noul#primitives#confidence
Read article
A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.
Jev Practices & Patterns#score#primitives#scoring
Read article
From the Jev community
Posts from builders shipping with Jev right now.