svrnmesh/corpus-engine-capstone

The Federalist Papers (capstone)

Hamilton, Madison and Jay's 85 essays (Project Gutenberg #1404), public domain, as the capstone's open-layer corpus.

What this is

A corpus published by corpus-engine as open Parquet: The Federalist Papers (capstone) documents, 1624 chunks and 1624 vectors in 10 file(s), 6177 rows. The search index is built from these layers and is not part of the repository.

Layers

layer	rows	bytes
documents	85	437564
chunks	1624	597938
embeddings	1624	3130799
ontology	2844	317587

Text bytes: 1164068. Published corpus footprint: 3.579x (corpus bytes ÷ text bytes). Ontology layer: 317587 bytes (0.273x of the text); reported separately, outside the corpus footprint.

Sources and licence

• Recipe: federalist-capstone (The Federalist Papers (capstone)), sha256 c04baef6668db6a70cf16b83b3d53d46cd99ab0b708666a94cbf16ad42b2e045.
• Licence: Public Domain (card id pddl).
• Source: {"type":"local_file","path":"/Users/alexsbryan/dev/corpus-engine-core/quality/capstone/federalist.jsonl"}.
• Custody grant: public-web.

Embed space

• Model: qwen-embedding-0.6b, 1024 dimensions, stored as f16.
• Pooling: last; normalization: application.
• Similarity: cosine.
• Query prompt:

Instruct: Given a web search query, retrieve relevant passages that answer the query
Query: 

• Document prompt: none
• EOS: <|endoftext|>

Chunk text

A chunk's stored text is title_headed(title, document_text[start_char..end_char]):

if title is present and the slice does not start with it:
    title + "\n\n" + slice
else:
    slice

Chunks

One row per chunk: chunk_id, doc_id, ordinal, start_char, end_char, text, content_hash, embed_hash. Join documents by (doc_id, ordinal); the span is in Unicode code points.

Embeddings

One row per distinct embedded text: embed_hash is BLAKE3 of the exact text the embedder saw, and embedding is the fixed-size vector in the embed space above.

Ontology — atlas

The pass's records, derived from the evidence and keyed to it:

{
  "ontology_version": "2.5",
  "pass": "atlas",
  "pass_version": "0.1.0",
  "config_sha256": "101c4dbe228ad4095074126b4a796b2c5e66c5ae40ae8c41cc13d9fe9d76b7d1",
  "model": "qwen-embedding-0.6b",
  "input_fingerprint": "1b355b45b171641fbf0a8d8caf73fbb1ea40c7f553b7571331c662695d806287"
}

Each row cites doc_id/ordinal/chunk_id/quote and carries its input_hash; the records store is rebuilt from this family.

Read it

From the repository root, DuckDB:

SELECT doc_id, title, url, text FROM read_parquet('documents/*.parquet') LIMIT 5;
SELECT chunk_id, doc_id, ordinal, start_char, end_char, text FROM read_parquet('chunks/*.parquet') LIMIT 5;
SELECT embed_hash, embedding::FLOAT[1024] FROM read_parquet('embeddings/*.parquet') LIMIT 5;
SELECT record_id, record_type, kind, chunk_id, quote FROM read_parquet('ontology/atlas/*.parquet') LIMIT 5;

datasets:

from datasets import load_dataset
documents = load_dataset("svrnmesh/corpus-engine-capstone", "documents", split="train")
chunks    = load_dataset("svrnmesh/corpus-engine-capstone", "chunks", split="train")
vectors   = load_dataset("svrnmesh/corpus-engine-capstone", "embeddings", split="train")
records_atlas = load_dataset("svrnmesh/corpus-engine-capstone", "ontology-atlas", split="train")

Optional snapshot

federalist-capstone.tar.zst (62020183 bytes, sha256 f8d7ce5b86cfd1914a30dba8280c88b60604446d95ef8a711976475197a061f8) is the tar.zst index shortcut; the open layers above are the primary and are rebuilt without it.

Install with CEC

cec-enrich install recipe.toml --embed-model qwen-embedding-0.6b