Upload PROVENANCE.md with huggingface_hub

PROVENANCE.md

@@ -0,0 +1,153 @@
+# Data Provenance
+
+This document describes the origin, processing method, and lineage of every table in the Epstein Document Archive.
+
+## Overview
+
+The dataset combines two OCR pipelines and multiple community-curated layers:
+
+| Source | Files | Method | Quality |
+|--------|-------|--------|---------|
+| Gemini OCR | 848,228 | Gemini 2.5 Flash Lite | High (structured JSON output) |
+| Community Tesseract | 537,622 | Tesseract OCR via community repos | Variable |
+| Community curation | -- | Manual + automated analysis | Curated |
+| ML recovery | 39,588 pages | Redaction recovery model | Experimental |
+
+**Success rate: 99.97%** (1,385,850 / 1,386,322 attempted).
+
+## Per-table provenance
+
+### `documents` (1,413,765 rows)
+
+| Subset | Rows | Source | Method |
+|--------|------|--------|--------|
+| `ocr_source IS NULL` | 848,228 | DOJ FOIA PDFs | Gemini 2.5 Flash Lite (`gemini-2.5-flash-lite`) via Google AI API. Each PDF page sent as image, returns structured JSON with full text, document type, date, entity extraction, and metadata. |
+| `ocr_source = 'tesseract-community'` | 537,622 | Community repositories | Tesseract OCR. Primary source: [rhowardstone/Epstein-research-data](https://github.com/rhowardstone/Epstein-research-data) for DataSet 9 (531K files). Remaining community files fill gaps in DataSets 2-5, 12, FBIVault, and HouseOversightEstate. |
+
+**Identifying OCR source**: Check the `ocr_source` column. NULL = Gemini, `"tesseract-community"` = community Tesseract.
+
+### `entities` (8,542,849 rows)
+
+Extracted entities (people, organizations, locations, dates, reference numbers) linked to their parent document via `file_key`.
+
+- **Gemini documents**: Entities extracted by Gemini as part of OCR (structured JSON output includes entity arrays).
+- **Community documents**: Entities extracted by post-processing NER on Tesseract OCR text.
+- **Entity types**: `person`, `organization`, `location`, `date`, `reference_number`.
+
+### `chunks` (2,039,205 rows)
+
+Text chunks derived from `documents.full_text` using ~800-token splitting with overlap. Used for retrieval-augmented generation (RAG).
+
+- **Source**: Derived from the full text of each document.
+- **Method**: Token-count-based splitting with character offset tracking.
+- **Fields**: `file_key`, `chunk_index`, `content`, `token_count`, `char_start`, `char_end`.
+
+### `embeddings_chunk` (1,958,052 rows)
+
+768-dimensional float32 embedding vectors for text chunks.
+
+- **Model**: `gemini-embedding-001` (Gemini Embedding v1).
+- **Input**: `chunks.content` (the chunk text).
+- **Coverage**: 96% of chunks (1,958,052 / 2,039,205).
+- **Format**: In Parquet, stored as `list<float32>` with fixed size 768. In SQLite, stored as raw `float32` BLOB (3,072 bytes per embedding).
+
+### `embeddings_summary` (1,413,508 rows)
+
+768-dimensional float32 embedding vectors for document summaries.
+
+- **Model**: `gemini-embedding-001` (Gemini Embedding v1).
+- **Input**: First ~2,000 characters of `documents.full_text`.
+- **Coverage**: 99.98% of documents (1,413,508 / 1,413,765).
+- **Format**: Same as chunk embeddings.
+
+### `persons` (1,614 rows)
+
+Curated registry of people mentioned across the corpus.
+
+- **Source**: Community curation, combining automated entity extraction with manual verification.
+- **Fields**: `canonical_name`, `slug`, `category` (perpetrator/victim/associate/other), `aliases` (JSON array), `search_terms`, `sources`, `notes`.
+- **Upstream**: `release/persons_registry.json`.
+
+### `kg_entities` (467 rows)
+
+Knowledge graph entities with metadata.
+
+- **Source**: Community curation -- manually identified key entities with occupation, legal status, and mention counts.
+- **Fields**: `name`, `entity_type`, `description`, `metadata` (JSON with occupation, legal_status, mention counts).
+- **Upstream**: `release/knowledge_graph_entities.json`.
+
+### `kg_relationships` (4,190 rows)
+
+Knowledge graph relationships between entities.
+
+- **Source**: Community curation -- relationships extracted from document analysis.
+- **Fields**: `source_name`, `target_name`, `relationship_type` (e.g., traveled_with, associated_with, employed_by), `weight`, `evidence`, `metadata`.
+- **Relationship types**: traveled_with, associated_with, employed_by, legal_representative, financial_connection, and others.
+- **Upstream**: `release/knowledge_graph_relationships.json`.
+
+### `recovered_redactions` (39,588 rows)
+
+Text recovered from behind redaction bars in DOJ documents.
+
+- **Source**: Community ML analysis of redacted document pages.
+- **Method**: Machine learning model trained to reconstruct text obscured by redaction bars.
+- **Fields**: `file_key`, `page_number`, `reconstructed_text`, `interest_score`, `names_found`, `document_type`.
+- **Quality**: Experimental. Higher `interest_score` indicates more significant recovered content.
+- **Upstream**: `release/redacted_text_recovered.json.gz`.
+
+### `provenance` (pipeline audit trail)
+
+Three sub-tables from the OCR processing pipeline:
+
+| Sub-table | Rows | Description |
+|-----------|------|-------------|
+| `provenance/files` | 1,386,322 | Per-file processing record: SHA-256 hashes, status, tokens, latency, model |
+| `provenance/audit_log` | ~3.6M | Append-only event log: every processing step recorded |
+| `provenance/runs` | 120 | Pipeline run metadata: start/end times, success/failure counts |
+
+- **Source**: `epstein_provenance.db` -- the forensic audit trail from the OCR pipeline.
+- **Key fields on `files`**: `file_key`, `pdf_sha256`, `output_sha256`, `status`, `input_tokens`, `output_tokens`, `api_latency_ms`, `model_used`.
+
+## Dataset-level breakdown
+
+### DOJ FOIA Releases
+
+| Dataset | Total Files | Gemini | Community | Description |
+|---------|-------------|--------|-----------|-------------|
+| DataSet 1 | 3,158 | 3,158 | 0 | Initial release |
+| DataSet 2 | 574 | 49 | 525 | |
+| DataSet 3 | 67 | 49 | 18 | |
+| DataSet 4 | 152 | 49 | 103 | |
+| DataSet 5 | 120 | 49 | 71 | |
+| DataSet 6 | 13 | 13 | 0 | |
+| DataSet 7 | 17 | 17 | 0 | |
+| DataSet 8 | 10,595 | 10,595 | 0 | |
+| DataSet 9 | 531,279 | 0 | 531,279 | Entirely community-processed |
+| DataSet 10 | 503,154 | 502,548 | 606 | |
+| DataSet 11 | 331,655 | 331,651 | 4 | |
+| DataSet 12 | 152 | 50 | 102 | |
+
+### Non-DOJ Sources
+
+| Dataset | Files | Source | OCR |
+|---------|-------|--------|-----|
+| FBIVault | 22 | FBI Vault FOIA | Community Tesseract |
+| HouseOversightEstate | 4,892 | House Oversight Committee | Community Tesseract |
+
+## Processing failures
+
+472 source PDFs could not be processed. These are documented in `release/epstein_problems.json` with:
+- `file_key` -- EFTA identifier
+- `doj_url` -- Original DOJ download URL
+- `error_message` -- Why processing failed
+- `category` -- Failure type (empty_source_pdf, corrupt_source_pdf, api_disconnect, doj_file_unavailable)
+
+## Reproducibility
+
+The full processing pipeline is available at [github.com/kevinnbass/epstein](https://github.com/kevinnbass/epstein):
+- `process_pdfs_gemini.py` -- Gemini OCR pipeline
+- `import_community.py` -- Community data importer
+- `validate_dataset.py` -- 14-check validation suite
+- `epstein_audit.db` -- Complete audit trail with per-file SHA-256 checksums
+
+Every output JSON has a SHA-256 checksum recorded in the provenance database, enabling verification that published data matches pipeline output.