DeepSearcher stopping-policy experiment
Experiment using Jev Noul judgments on accumulated evidence to decide when Zilliz's DeepSearcher agent should stop searching.
Overview
DeepSearcher is Zilliz's open-source deep-research tool that combines LLMs with vector databases such as Milvus to search, evaluate and reason over private data and produce answers and reports. It targets enterprise knowledge management, Q&A and retrieval.
The directory entry concerns a standalone evaluation built on it. According to the list summaries, it uses Jev Noul judgments on the accumulated evidence to decide whether the agent should stop or continue searching within a search-round budget, and compares stopping behaviour, evidence recall and decision cost. The main project README does not describe the Jev experiment, so details of the setup and results come only from the listings. DeepSearcher itself installs from PyPI, supports optional extras such as Ollama, and recommends uv for development installs.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is the evidence gathered so far sufficient to stop searching?(inferred) | yes · no |
How it works
- Agent runs a search round and accumulates evidence
- Jev judges whether evidence is sufficient
- Stop and answer, or continue until the round budget is exhausted
- Compare stopping behaviour, recall and cost against baselines
Strengths
- Cheap stopping decision separate from the LLM
- Measures recall and cost trade-offs
Caveats
- Jev experiment is not described in the main README
- Results not reviewed here
Getting started
pip install deepsearcher (Python 3.10 recommended); see the repository for the stopping-policy evaluation.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- library / CLI
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Agentic search: a standalone evaluation uses Jev Noul judgments on accumulated evidence to decide whether to stop or continue within a search-round budget, comparing stopping behavior, evidence recall, and decision cost.
yibie/awesome-jevOpen Source Deep Research alternative with Jev search-stopping evaluation.
logicrw/awesome-jev-projects
Frequently asked questions
What does Jev decide?
Whether the research agent has enough evidence to stop searching.
Is DeepSearcher itself a Jev project?
No, it is a general deep-research tool; the Jev part is a stopping-policy evaluation.
What is measured?
Stopping behaviour, evidence recall and decision cost, per the summaries.
Related projects
Python agent runtime where Jev Choice decisions pick tools and arguments, and low-confidence picks go to an LLM for review.
ServicebetaPythonJIT-JEV Context OS: deterministic context compiler plus a Jev System 1 gate that scores AST relevance, test proofs and tool targets.
BenchmarkresearchPythonCraftax survival agent where Jev picks one Choice per step from code-listed actions, with optional LLM goals, plus ablations.
ProjectresearchPythonDeepSeek Harness plugin with jev_decision tools, an opt-in risk and loop auto-guard, and a live self-test benchmark against the Jev API.
BenchmarkbetaTypeScriptPre-registered benchmark where Jev makes every search and evidence decision for a 2B local model answering SimpleQA questions.
BenchmarkresearchPythonBenchmark where Jev steers an agent through 100 key-and-door grid puzzles, one Choice per move, scored against an exact BFS solver.
CLIresearchPython