jev-ultralightspeed
Python library that packs many items into each Jev request for bulk yes/no classification and raises on any unanswered item.
Overview
jev-ultralightspeed answers one question about each item in a large pile of text (support tickets, reviews, log lines, a table column) using TypeSafe's Jev. A single classify(items, question) call returns answers you can filter on, for example tickets that need a human today. It packs up to 32 items into one request; if any item in a pack comes back unanswered it raises and names the item instead of returning a partial result.
The README is careful about its numbers. Packing gives 32 times the throughput under a ceiling counted in requests, which it calls arithmetic rather than measurement; the measured result over 30,000 judgments against human labels is 41% lower cost with no detectable accuracy difference. On a synthetic million-message corpus it estimates half an hour and $4.99 packed versus seventeen hours and $14.97 one per item. An optional guidance="once" mode stops repeating the question per item, cutting a 32-item request from 4,598 to 1,777 billed tokens at about 0.2 points of agreement. It is Apache-2.0 on PyPI with no required dependencies and CI tests.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | User-supplied yes/no question per item, e.g. Does this message need a human to act on it today? | — |
How it works
- Pass a list of items and one question
- Pack items into requests of up to 32
- Validate every item was answered; raise and name any missing item
- Return per-item answers to filter in code
Strengths
- Measured 41% cost saving over 30,000 labelled judgments
- Throughput gain under request-based rate limits
- Fails loudly on missing answers
- No required dependencies
Caveats
- Throughput gain depends on request-counted rate limits
- guidance=once trades a little agreement for fewer tokens
- Savings depend on item length
Getting started
pip install jev-ultralightspeed, then `from jev_ultralightspeed import classify`.At a glance
- Type
- Demo
- Status
- beta
- Language
- Python
- Runtime
- Library
- Has code
- yes
- Has tests
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
BRRRRRRRRRRRRRRRRRRRRRR.
hellogumbo/awesome-jevPacks many items into one Jev request for bulk classification. If any item in a pack comes back unanswered it raises and names the item rather than returning a partial result.
MrJev/awesome-jev
Frequently asked questions
How much does packing save?
The README measures 41% lower cost with no detectable accuracy difference over 30,000 judgments.
What happens if an item is not answered?
The call raises and names the missing item instead of returning a partial result.
Is the 32x throughput guaranteed?
No, it holds only while rate limits count requests rather than tokens.
Related projects
Hack Club's lightweight LLM proxy, listed for adding a Jev forwarding endpoint behind its auth, limits and usage logging.
MCP serverproductionTypeScriptnpm library that rate-limits, spend-caps, caches and redacts every Jev request made in a loop.
LibrarybetaJavaScriptPython and Node library that classifies Hugging Face dataset rows on typed Choice, Score and Noul dimensions using Jev.
ProjectbetaPython, TypeScriptGo gateway that pools multiple TypeSafe accounts behind the native System One API with user keys, quotas and failover.
ServicebetaGoMoves a coding agent's bulk semantic judgments (triage, tagging, routing) to Jev; an A/B reports 78% fewer tokens at similar accuracy.
ProjectbetaPythonRust CLI for Jev that lints typed question sets offline, then sends them and prints probabilities, labels and scores as JSON.
CLIbetaRust