應用研究／科學資料工程
Astro-COLIBRI Parses Astronomy Circulars with a Constrained LLM, Turning a Decade of Free Text into 68,393 Observations
A new pipeline converts photometry, redshift, and observation times from NASA GCN Circulars into event-level structured data and is already operating in real time. An internal audit recorded 99.80% field-level accuracy, but time interpretation and observatory attribution remain the main sources of error, so researchers must consult the original text.
The Astro-COLIBRI team has unveiled an operational LLM data pipeline for processing NASA GCN Circulars published by observatories as free text. A continuously running listener first receives each circular, associates it with transient events such as gamma-ray bursts, and performs deterministic pre-analysis. It then calls a large language model constrained by a fixed schema to extract photometric values, filters, times, upper limits, redshifts, contact information, and observing facilities. After unit and time normalization, the results are written to the Astro-COLIBRI API for use by its website, mobile app, CSV and VOTable exports, and optical afterglow plots.
According to the paper, after tuning, the workflow successfully processed all 1,775 circulars evaluated in the first half of 2026, with no workflow-level failures. Researchers manually reviewed 210 Circulars and confirmed 25,827 of 25,880 unambiguously assessable fields, yielding field-level accuracy of 99.80%. The 53 errors were concentrated in eight circulars and involved time interpretation and facility attribution. In a comparison with the independent GRBweb dataset, redshift values agreed for 228 of the 249 events shared by both datasets. The remaining discrepancies arose from upper limits, candidate-host estimates, or later revisions rather than simple numerical transcription errors.
The team also applied the pipeline to the complete archive dating back to 2016, organizing 68,393 observations for 5,787 events from 26,811 reports. It now processes new circulars in real time. The reusable components have been released as the Python package `astro-colibri-circular-parser`, which includes deterministic pre-analysis, schema-constrained extraction, event association, data enrichment, and consistency checks. This application deserves more attention than conventional document summarization: model output is bounded by a data contract and post-processing, and is validated through manual field-level audits and cross-checks against an external catalog. However, the sample comes from a single specialized writing style, and the audit was conducted internally by the team. Time expressions, non-standard filters, and Vega/AB calibration conversions may still introduce scientific bias. Any formal analysis should retain the Circular ID and consult the original circular rather than treating high average accuracy as a guarantee for every individual record.