Featured · Cohere Labs Community blog · November 2026
How FeatureGraph builds reproducible summaries of time-series
scientific and engineering signals, and how its compiler and
language-model query layer fit together.
Research updates
A public record of how FeatureGraph is evolving, including positive
and negative results, as the project develops explicit, inspectable
behavioral representations.
November 2026
A post on the Cohere Labs Community blog introducing FeatureGraph:
how it builds reproducible summaries of time-series scientific and
engineering signals, and how its compiler and language-model query
layer fit together.
October 2026
An LLM writes a plain-English question as a request in a small,
fixed query language. A separate program checks it, runs it on
the table and saves a receipt. The page shows 13 recorded runs
on a small made-up table. It is a demonstration, not a
measurement of accuracy.
August 2026
The public from_state_sequence adapter materialized
20,700 unchanged CLaP labels as nine bounded objects and eight
temporal relations. All eleven study checks passed, including
exact reconstruction of the original state sequence.
August 2026
On subject 1, all 169 reference candidate breaths matched
FeatureGraph results. Across subjects, the same fixed rule
over-segmented, and robust MAD scaling increased recall while
producing still more extra candidates. The study supports
FeatureGraph as a deterministic preservation and audit layer, not
an autonomous judge of meaningful peaks.
August 2026
Waveform area relative to an explicit baseline is recorded with
parent identity, completeness, timing, rates, and centroid. It is
not interpreted as physical inhaled volume without sensor
calibration.
July 2026
Rising, falling, and inactive behavior is represented as explicit
Transition objects. This development line now motivates the
transition-only successor study.
July 2026
A summary of the research, architecture, testing, reproducibility
work, and public release that took FeatureGraph from an initial
idea to version 0.1.0a1.
July 2026
Why accumulated quantities should be represented explicitly and
what the object exposes beyond a cumulative sum.
June 2026
The original motivation: convert observation sequences into
explicit structures that can be inspected, measured, related, and
queried without repeatedly rebuilding the same logic.
Using FeatureGraph on your own data?
Have an analysis built on your time-series data, or build your own
with the open-source compiler.
FeatureGraph for Teams →