Semantic Condensation of High-Cardinality Time Series for LLM-Driven Observability
Abstract
Large language models do not usually fail because they cannot reason. They fail because of the inputs they are given, and in observability that means raw, high-cardinality, high-volume time series data that carries very little signal per token. This paper shows how to turn that raw data into structured, high-signal input, and demonstrates the improvement in outcomes that follows from doing so. The underlying argument generalizes well beyond observability: in any system where the question is extracting signal, condensing the input matters more than enlarging the context window, and the instinct to supply more is usually the wrong response to a disappointing result. The paper treats input preparation as the engineering problem rather than the preliminary to it. Readers take away a different way of preparing data for a language model, oriented toward the signal the model actually needs rather than the volume it can technically accept.
- Published in
- Proceedings of the Raptors Conference, Vol. 1 (2026)
- Conference
- Raptors Conference 2026, 21 October 2026, online
- Accepted
- 2026-09-21
- DOI
- https://doi.org/10.68495/rc.2026.C063
- Presentation
- Slides (PDF)
- Licence
- © 2026Akila Balasubramanian. Published under Creative Commons Attribution 4.0 International (CC BY 4.0).
- Publisher
- Hackathon Raptors, Community Interest Company 15557917, London, United Kingdom
- Permanent address
- https://proceedings.raptors.dev/balasubramanian-semantic-condensation-high-cardinality
Open access. Every entry in the Proceedings is free to read, with no charge to authors and no charge to readers. Slides are served from this site, so they stay available at the paper's permanent address.
Cite this paper
Akila Balasubramanian (2026). Semantic Condensation of High-Cardinality Time Series for LLM-Driven Observability. Proceedings of the Raptors Conference, Vol. 1 (2026). Hackathon Raptors, London. https://doi.org/10.68495/rc.2026.C063
@inproceedings{raptors2026-c063,
title = {{Semantic Condensation of High-Cardinality Time Series for LLM-Driven Observability}},
author = {Balasubramanian, Akila},
booktitle = {Proceedings of the Raptors Conference},
volume = {1},
year = {2026},
publisher = {Hackathon Raptors},
address = {London, United Kingdom},
doi = {10.68495/rc.2026.C063},
url = {https://doi.org/10.68495/rc.2026.C063},
abstract = {Large language models do not usually fail because they cannot reason. They fail because of the inputs they are given, and in observability that means raw, high-cardinality, high-volume time series data that carries very little signal per token. This paper shows how to turn that raw data into structured, high-signal input, and demonstrates the improvement in outcomes that follows from doing so. The underlying argument generalizes well beyond observability: in any system where the question is extracting signal, condensing the input matters more than enlarging the context window, and the instinct to supply more is usually the wrong response to a disappointing result. The paper treats input preparation as the engineering problem rather than the preliminary to it. Readers take away a different way of preparing data for a language model, oriented toward the signal the model actually needs rather than the volume it can technically accept.}
}
TY - CONF TI - Semantic Condensation of High-Cardinality Time Series for LLM-Driven Observability AU - Balasubramanian, Akila T2 - Proceedings of the Raptors Conference VL - 1 PY - 2026 DA - 2026/09/21 PB - Hackathon Raptors CY - London, United Kingdom DO - 10.68495/rc.2026.C063 UR - https://doi.org/10.68495/rc.2026.C063 AB - Large language models do not usually fail because they cannot reason. They fail because of the inputs they are given, and in observability that means raw, high-cardinality, high-volume time series data that carries very little signal per token. This paper shows how to turn that raw data into structured, high-signal input, and demonstrates the improvement in outcomes that follows from doing so. The underlying argument generalizes well beyond observability: in any system where the question is extracting signal, condensing the input matters more than enlarging the context window, and the instinct to supply more is usually the wrong response to a disappointing result. The paper treats input preparation as the engineering problem rather than the preliminary to it. Readers take away a different way of preparing data for a language model, oriented toward the signal the model actually needs rather than the volume it can technically accept. LA - en ER -