Grepstein

Grepstein v1#

Grepstein is the next version of the earlier Epstein Search example. The v1 implementation targets the smaller DOJ releases: datasets 1–7 and 12. Datasets 8–11 are deferred. An existing court PDF collection can also be imported.

The implementation is tracked in Colony #818. Availability and dataset coverage depend on the Cloud deployment and its qualification run; this page does not assert that every release is already indexed.

One extraction, several search views#

Original PDFs and prepared PDF windows live in private GCS. Antfly fetches those remote PDFs and owns native text extraction and optional Vertex OCR. The example has no DocsAF dependency.

Full-text search comes first. Embeddings and GLiNER relationship extraction can be added afterward using the existing native artifacts. Each result retains source and page evidence. Graph neighborhoods expose extracted relationships and path evidence for inspection; model outputs require verification against the underlying records.

Canonical grep items copy the extracted page text without rewriting it. Matches are exact within each page. Cross-page multiline matching is outside the v1 contract. Colony's existing scheduled GCS source jobs publish these text projections through scoped LinearMerge and the ordinary Antfly source catalog.

Explore and operate#

The hosted explorer keeps Antfly and GCS credentials on the server and exposes fixed read-only queries. It supports full-text, semantic, and graph evidence modes, with links to original PDFs.

The dedicated antfly-showcases infrastructure keeps originals and derived data private. The database remains an ordinary managed Antfly Cloud instance, using existing resource configuration rather than a separate control plane or a special showcase tier.

See the Grepstein example for acquisition, publication, indexing, source jobs, and hosting instructions. The historical showcase URL is retained so existing links continue to resolve.