README.md

metadata

license: other
pretty_name: AGGS GitHub structural corpus
tags:
  - agent-guided-graph-search
  - github
  - code-search
  - dependency-graph
  - bnbcode
task_categories:
  - other
size_categories:
  - 10M<n<100M

AGGS GitHub structural corpus

This snapshot supports repository discovery and dependency-graph navigation without redistributing source content. Its manifest records the exact build ID, schema version, artifact hashes, byte sizes, row counts, and audited structural symbol count.

Contents

• canonical GitHub owner/repository and PyPI package identifiers;
• typed package/repository/native dependency relations and numeric evidence;
• repository-relative source paths;
• qualified symbol names and symbol kinds; and
• whitespace-normalized signatures capped at 1,024 UTF-8 bytes.

The graph is a PostgreSQL custom-format dump compressed with xz. The matching Tantivy search index is a zstd-compressed tar archive. A target-specific Rust query executable is the third artifact. The graph and index carry one build ID and graph schema version; mixing generations is rejected, and all three artifacts are SHA-256 verified.

Deliberate exclusions

The snapshot contains no source archives, source bodies or snippets, comments, docstrings, README text, repository/package descriptions, copied license text, or local archive paths. Archive availability is false. Consequently, source-search and fetch are unavailable; use ordinary search for structural facts and obtain source separately under its repository's terms when needed.

Sources and licensing boundary

The catalog combines generated dependency data from deps.dev's public BigQuery tables, PyPI package metadata, GitHub public-repository metadata, and structural facts extracted from public repositories. deps.dev documents its generated data under CC BY 4.0; that license does not relicense fields originating from PyPI, GitHub, or repositories. GitHub public visibility likewise does not place repository content in the public domain.

The snapshot therefore does not claim one blanket copyright license over all identifiers or declarations. Its paths, names, kinds, and signatures are provided as structural/API facts under the applicable source terms and law. The signature allowance is a product risk decision, not legal advice or a conclusion for every jurisdiction. Users who retrieve source must comply with the corresponding repository license.

Primary provenance references:

• https://docs.deps.dev/faq/
• https://docs.deps.dev/bigquery/v1/#data
• https://cloud.google.com/bigquery/public-data
• https://docs.github.com/en/site-policy/github-terms/github-terms-of-service

Verification and installation

The release pipeline first runs github_structure_audit.py over the five-field JSONL stream, then github_snapshot.py rejects incomplete or non-structural builds, prose/archive leakage, mixed identities, and count mismatches before it writes manifest.json. Install with:

bnbcode research install --code --execute

The installer verifies both SHA-256 hashes, graph row counts, build/profile identity, and archive paths/symlinks before atomically placing the index and writing ~/.config/opencode/arxiv_graph.env discovery settings.

Rebuilding from scratch

The builder/ directory holds the acquisition-time metadata inputs for recreating this corpus without BigQuery or the GitHub search API: the package/repository catalog, the PyPI dependency edges, and the PyPI package metadata export. These are public registry facts, not source content. aggs/scripts/github_build_from_scratch.py stages them, shallow-clones every catalogued public repository, extracts native dependency evidence, runs the closed-schema structural extractor and independent audit, and builds the same graph schema and structural profile as this snapshot. Because cloning observes current default-branch HEADs, a fresh build matches this snapshot's structure and pipeline but not necessarily its exact rows where repositories have moved or disappeared. Source is fetched directly from each repository under that repository's license.

Current snapshot

Snapshot date	Repositories	Packages	Graph edges	Artifact bytes	Backend target
2026-08-19	52,520	509,232	2,468,590	1,580,778,877	linux-x64

Build ID: 1787128807-160c21c9d3c3dc41. See manifest.json for artifact hashes, individual byte sizes, table counts, and graph schema version.

Builder inputs

Repositories	PyPI edges	Metadata rows	Builder bytes
47,272	2,250,939	38,542	367,477,631

The builder/ directory contains the repository catalog and PyPI dependency exports needed to rebuild this corpus from scratch by cloning the catalogued repositories. See builder/manifest.json for hashes and the dataset card body for the rebuild procedure.