# GraphRAG remote release schema

`manifest.json` declares `primary_key` (`entry_cid`), BM25 constants, vector
layout, and compact index descriptors. Verify `sha256` / `cid` before parsing
a shard.

## Identity

- `entry_cid`: CIDv1 primary identity.
- `document_index`: dense `0..N-1` pointer used only inside postings.
- Data shards are Zstandard Parquet with at most 4,096 rows.

## BM25

`indexes/bm25_keyword_shards.parquet` maps inclusive `first_key`/`last_key`
term ranges to posting shards. Nested posting rows contain `term` plus aligned
`document_indices`, `document_lengths`, `title_frequencies`, `body_frequencies`,
and `idf`. Posting shards must be globally sorted by term with disjoint
ranges; a term never spans two shards. In-memory cells: `write_term_sorted_posting_shards`.
Already-dumped shards (any consumer): `pack_range_routed_family` at build time
or `repair_graphrag_range_routing` on a release root. Both use the same
resource-aware process pool. Exploded `(term, legal_id, tf)` shards remain readable.

## Vectors

`indexes/vector_chunks.parquet` maps centroid groups to at most two physical
shards of 4,096 rows. Probe a few centroids, then exact-score inside those
shards.

## Graph

`indexes/graph_outgoing_adjacency.parquet` and
`indexes/graph_incoming_adjacency.parquet` map `node_cid` ranges to paged
adjacency. Each page holds at most 4,096 pointers. Similarity neighbors cannot
establish legal authority.