Unified data platform
for physical AI

Powering agentic systems with millisecond queries at petabyte scale and billions of records.

Autonomous drivingDrivingWarehouse robotsWarehouseManipulationManipulation

Search: construction zones where v42 disengagd

Compiled in 4 ms

near   embedding("construction zone")
where  software = 'v42'
  and  disengaged = true
limit  4

4 nearest matches

1d = 0.214in view
2d = 0.229
3d = 0.241
4d = 0.256

Search: palets in the aisles at reno-2 this week

Compiled in 5 ms

near   embedding("pallet in the aisle")
where  site = 'reno-2'
  and  ts >= date_trunc('week', now())
limit  4

4 nearest matches

1d = 0.197in view
2d = 0.213
3d = 0.224
4d = 0.238

Search: frnaka arms failing to grasp clear cups

Compiled in 3 ms

near   embedding("grasping a clear cup")
where  robot = 'franka'
  and  success = false
limit  4

4 nearest matches

1d = 0.188in view
2d = 0.205
3d = 0.219
4d = 0.233

Simulated roof lidar. Simulated AMR lidar. Simulated depth camera. Illustration. The quoted text is embedded with the same model as the scans; matches are the nearest rows that pass the filter.

What we can do

Fig. 1 A failure's nearest neighbors in embedding space. Rows that fail the filter drop out.
Triage a failure
Start from a disengagement or failed grasp and find every event like it on the same platform, software, and conditions, and send the set to labeling.
```
near   clip(:event)
where  platform = 'a'
  and  software = 'v42'
```
Fig. 2 A hit on camera, lidar, and CAN streams, widened to two seconds before and five after.
Pull the context
Expand every hit to the seconds around it and hand the source references to the viewer you already use.
```
near   clip(:hit)
expand window 2 s before, 5 s after
```
Fig. 3 The same six situations: v41 handled five, v42 missed three.
Catch a regression
Search the same situations under two releases and compare what each one did before you ship.
```
near   clip(:miss)
where  software in ('v41', 'v42')
```
Fig. 4 How often each scenario occurs. The matches sit in the rare end.
Mine the long tail
Describe a rare scenario in words and get the closest real examples, even when the conditions leave one row in a million.
```
near   embedding("unprotected left, heavy rain")
where  time_of_day = 'night'
```
Fig. 5 The best one or two frames per episode, balanced across three cities.
Build an eval set
Assemble test sets that cover the scenarios you care about: the best few per episode, spread across cities, weather, and sites.
```
near   clip(:scenario)
where  city in ('sf', 'phx', 'aus')
top    5 per episode_id
```
Fig. 6 Teleop trajectories; the three closest to the new task are highlighted.
Pick demonstrations
Find the teleop episodes closest to a new task, on the right robot and gripper, to fine-tune a policy.
```
near   embedding("insert peg into hole")
where  robot = 'franka'
  and  success = true
```

Illustrations. Each figure is drawn from its query, not from a measurement.

Works on the data you already have.

Bring your datasets and logs
FiftyOne or LeRobot datasets, MCAP or rosbag logs. Falcon keeps an embedding per frame, clip, or episode, the fields you filter on, and a reference back to the source.
Ask with meaning and conditions
Start from an example or a sentence. Filter on any column and time range, and group results by episode or run.
Get exactly what matches
Every result passes your filter, and you get the k you asked for whenever that many exist. EXPLAIN SEARCH shows the plan that ran.

The filter decides the plan.

Falcon estimates how many rows pass, and where they sit, picks a plan, and switches when the estimate was wrong.

time_of_day = 'day'Broadweather = 'rain' and speed_mps > 5Narrow, near the queryintervention and software in ('v41', 'v42')Tinyroad = 'construction'Far from the query

Illustration. A 2-D stand-in for embedding space: 900 rows whose columns correlate with their neighborhoods. Every number here comes from running each plan on it.

Plan

Graph traversal

Rows that pass: 75%
Near the query: 72%

Most rows pass, so walking the similarity graph toward the query finds matches almost immediately.

EXPLAIN SEARCH

rows pass: 75% (675 of 900)
plan: graph traversal, ef = 24
visited: 45 rows, 33 pass
result: k = 5 d 0.003–0.015 recall 5/5

Plan

Partition search

Rows that pass: 13%
Near the query: 60%

13% pass, and they cluster around the query. Probing the three closest partitions covers them.

EXPLAIN SEARCH

rows pass: 13% (115 of 900)
plan: partition search, probe 3 of 16
scanned: 181 rows, 71 pass
result: k = 5 d 0.003–0.019 recall 5/5

Plan

Exact scan

Rows that pass: 0.6%
Near the query: 0%

Only 5 rows pass. Scoring every one is the fastest plan, and it’s exact.

EXPLAIN SEARCH

rows pass: 0.6% (5 of 900)
plan: exact scan of rows that pass
scored: 5 rows
result: k = 5 d 0.192–0.319 recall 5/5

Plan

Partition search, then exact scan

Rows that pass: 7.8%
Near the query: 0%

7.8% pass, but they sit far from the query. The probe found 1 of the 5 needed, so the planner switched to scoring all 70.

EXPLAIN SEARCH

rows pass: 7.8% (70 of 900)
plan: partition search, probe 3 of 16
scanned: 181 rows, 1 pass: too few
switch: exact scan of 70 rows that pass
result: k = 5 d 0.089–0.288 recall 5/5

Most fixed strategies have a cliff

Schematic, not a benchmark: textbook cost shapes with matches spread evenly. Correlation moves the cliffs, which is why the planner also watches yield while it runs. Published results will state recall, k, memory, and hardware.

Bring your hardest query.

Request early access

Unified data platform for physical AI

What we can do

Triage a failure

Pull the context

Catch a regression

Mine the long tail

Build an eval set

Pick demonstrations

Works on the data you already have.

Bring your datasets and logs

Ask with meaning and conditions

Get exactly what matches

The filter decides the plan.

Bring your hardest query.

Unified data platform
for physical AI