Benchmarks
Speed you can check.
Every number on this site, with the machine, the data and the method behind it.
6,304/spassages indexed per second, CPU only
1.57 Mpassages in one index, on one workstation
<10mstypical search on real workspaces (p50 5–9 ms, Oct 2026)
2 minto read 65,839 Wikipedia articles on a Mac with M5 Max
Measured: production measurement windows, Oct 2026 (search) · INDAGO test report v1.1, Apple M5 Max, no GPU (indexing) · website demo. How we measured →
Scale and speed
One workstation. No GPU.
| Measure | Result | Conditions |
|---|---|---|
| Indexing rate, sustained | 6,304 passages/s | Real patent full texts; ≈1.4 M documents per hour at 16.2 passages per document |
| Largest single index | 1.57 M passages | Indexed and sealed on one machine |
| Search on real workspaces, p50 / p95 | 5–9 / 9–22 ms | Measurement windows 5–11 Oct 2026, production builds, five real workspaces up to 1.9 M passages, measured while indexing ran |
| Search vs local RAG stacks, p50 | 7–20 ms vs 75–125 ms | Quiet run 30 Sep 2026 through the REST API, end to end; embeddinggemma, qwen3-embedding, hybrid RRF on the same machine |
| Time to searchable vs embedding RAG | 3.1 s vs 124 s | BEIR SciFact, 5,183 abstracts; qwen3-embedding took 269 s |
| Engine search at 920 k passages, p50 / p95 / p99 | 3.9 / 4.5 / 6.5 ms | 919,898 real EPO patent passages, isolated service, one worker; KPI run 24 Aug 2026, re-measured at 4.2–4.6 ms p50 |
| Throughput | 206 → 354 queries/s | Same index, 1 → 4 concurrent workers |
| First query after start, p50 | 32 ms | Same index, nothing cached yet |
| Open a 1 M-passage index | 0.5–1.3 s | Fresh index, 855 MB memory to mount; capacity study 2026-09-02 |
| Search while indexing, median / p99 | 54 / 168 ms | 1.17 M passages, measured during the TRL scale run with indexing running in parallel |
| Hard kill during indexing | 0 lost | Process killed at 807,894 passages; resumed from 802,812 with no committed data lost |
Answer quality
Fast and right.
| Benchmark | Score | What it tests |
|---|---|---|
| BEIR SciFact | 0.686 nDCG@10 | Finding the scientific evidence for a claim (BM25 reference 0.662) |
| Code symbols for agents | 0.98 / 0.89 nDCG@10 | Go / Python: finding the right symbol for a coding agent's question |
| RepoQA | 0.63 recall@1 | Finding the right function in real code from a description |
| ViDoRe | ≈0.75 nDCG@5 | Finding answers inside charts, scans and slides — CPU only |
Website demo: Wikipedia sample
| Measure | Result | Conditions |
|---|---|---|
| Articles | 65,839 | Deterministic 1% sample of English Wikipedia (hash rule), files in topic folders |
| Passages | 481,905 | After splitting into searchable passages |
| Time to searchable | 2 min | Apple M5 Max (Mac17,7), 128 GB, CPU only |
Hardware: Apple M5 Max, 18 CPU cores, 128 GB RAM, internal SSD plus external data volume. No accelerators. Source: INDAGO TRL 4/5 test report v1.1, October 2026 — download the PDF.
Find it in a blink.
Your documents stay on your Mac.
macOS · Apple silicon · Signed and notarised