Abstract
This disclosure describes techniques that leverage a large language model (LLM) to perform root cause analysis (RCA) on high-volume telemetry data. The techniques can handle datasets whose size exceeds the context window of the LLM by traversing the data using a stateful log-walker that recursively passes an investigation state between LLM inference calls. A vector search is executed to retrieve a domain-specific structured graph of expected state transitions. The log stream is ingested in sequential blocks. A mutable investigation state which is a structured object that tracks the validation status of the steps of the schema is maintained. The object is passed recursively between LLM inference calls. During a dynamic instruction injection phase, the validation results of block N programmatically alter the prompt instructions for block N+1, enabling the model to dynamically focus the attention mechanism on verifying specific causal precursors while discarding noise.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Phoughat, Aakash, "A Stateful, Hypothesis-Driven Framework for Long-Context Root Cause Discovery in Telemetry Analysis", Technical Disclosure Commons, (February 11, 2026)
https://www.tdcommons.org/dpubs_series/9322