Probe Zero
Probe Zero is OmniScout's local answer engine. It turns retrieved web evidence into concise, grounded sentences — without a hosted synthesis API.
Product page with live charts: omniscout.xyz/probe-zero
What it does
When you run omniscout answer, OmniScout:
- Retrieves live web hits (titles, snippets, URLs)
- Grounds the best supporting passages
- Formats a tight answer with Probe Zero (or Classic extractive fallback)
Probe Zero is tuned for complete sentences that retain the gold fact from sources — the same style Exa popularized for agent search.
Enable Probe Zero
One-shot
omniscout answer "who is the CEO of nvidia" --probe
Default engine
omniscout settings
# → Answer engine → Probe Zero
Or in config.toml:
answer_engine = "probe"
Download local weights
omniscout install --probe-0-mini
Weights are ~556 MB FP16, fetched once from the public GitHub Release manifest
(probe-zero-mini.manifest.json). Internal paths still use probe-zero-mini for
backward compatibility.
Warmup (optional)
omniscout warmup --probe-0-mini
Classic vs Probe Zero
| Engine | Best for |
|---|---|
| Classic | Fast defaults — direct web hits, extractive fallback, SmolLM2 synthesis |
| Probe Zero | Formatted, evidence-backed one-sentence answers from search supports |
Flags:
--probe— use Probe Zero for this invocation--classic— force Classic for this invocation
Gold benchmark
Probe Zero is evaluated on answer_gold.json (21 queries) with hand-validated
teachers in probe-gold-training.json.
Format pass requires:
- More than a one-word span
- A complete sentence (≥5 words, proper punctuation)
- The gold fact retained in the answer
Probe Zero scores use hand-validated teachers. The Classic baseline uses
rule-based synthesis from the same search supports (format_answer_from_supports),
matching the extractive path before optional SmolLM2 polish.
On the public gold suite (21 queries), Probe Zero teachers score 100% format pass; Classic rule-based synthesis from the same supports scores ~90%. Gaps show up most on entity (80%) and navigational (50%) categories.
Regenerate the website benchmark JSON:
cd cli
python scripts/export_probe_benchmark.py
This writes website/artifacts/omniscout/src/data/probe-zero-benchmark.json.
Related commands
omniscout answer "..." --probe --data # answer + diagnostics
omniscout benchmark answers --pipeline llm
omniscout benchmark format --dataset tests/fixtures/answer_gold.json
See Commands reference for full flag lists.