feat: Remove Document collection from schema

BREAKING CHANGE: Document collection removed from Weaviate schema

Architecture simplification:
- Removed Document collection (unused by Flask app)
- All metadata now in Work collection or file-based (chunks.json)
- Simplified from 4 collections to 3 (Work, Chunk_v2, Summary_v2)

Schema changes (schema.py):
- Removed create_document_collection() function
- Updated verify_schema() to expect 3 collections
- Updated display_schema() and print_summary()
- Updated documentation to reflect Chunk_v2/Summary_v2

Ingestion changes (weaviate_ingest.py):
- Removed ingest_document_metadata() function
- Removed ingest_document_collection parameter
- Updated IngestResult to use work_uuid instead of document_uuid
- Removed Document deletion from delete_document_chunks()
- Updated DeleteResult TypedDict

Type changes (types.py):
- WeaviateIngestResult: document_uuid → work_uuid

Documentation updates (.claude/CLAUDE.md):
- Updated schema diagram (4 → 3 collections)
- Removed Document references
- Updated to reflect manual GPU vectorization

Database changes:
- Deleted Document collection (13 objects)
- Deleted Chunk collection (0 objects, old schema)

Benefits:
- Simpler architecture (3 collections vs 4)
- No redundant data storage
- All metadata available via Work or file-based storage
- Reduced Weaviate memory footprint

Migration:
- See DOCUMENT_COLLECTION_ANALYSIS.md for detailed analysis
- See migrate_chunk_v2_to_none_vectorizer.py for vectorizer migration

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

generations/library_rag/.claude/CLAUDE.md

@@ -138,34 +138,30 @@ The core of the application is `utils/pdf_pipeline.py`, which orchestrates a 10-
 - `use_ocr_annotations=True` - OCR with annotations (3x cost, better TOC)
 - `ingest_to_weaviate=True` - Insert chunks into Weaviate
 
-### Weaviate Schema (4 Collections)
+### Weaviate Schema (3 Collections)
 
-Defined in `schema.py`, the database uses a normalized design with denormalized nested objects:
+Defined in `schema.py`, the database uses a denormalized design with nested objects:
 
 ```
 Work (no vectorizer)
   title, author, year, language, genre
-
-  ├─► Document (no vectorizer)
-  │     sourceId, edition, pages, toc, hierarchy
+  │
+  ├─► Chunk_v2 (manual GPU vectorization) ⭐ PRIMARY
+  │     text (VECTORIZED)
+  │     keywords (VECTORIZED)
+  │     workTitle, workAuthor, sectionPath, chapterTitle, unitType, orderIndex
   │     work: {title, author} (nested)
   │
-  │   ├─► Chunk (text2vec-transformers) ⭐ PRIMARY
-  │   │     text (VECTORIZED)
-  │   │     keywords (VECTORIZED)
-  │   │     sectionPath, chapterTitle, unitType, orderIndex
-  │   │     work: {title, author} (nested)
-  │   │     document: {sourceId, edition} (nested)
-  │   │
-  │   └─► Summary (text2vec-transformers)
-  │         text (VECTORIZED)
-  │         concepts (VECTORIZED)
-  │         sectionPath, title, level, chunksCount
-  │         document: {sourceId} (nested)
+  └─► Summary_v2 (manual GPU vectorization)
+        text (VECTORIZED)
+        concepts (VECTORIZED)
+        sectionPath, title, level, chunksCount
+        work: {title, author} (nested)
 ```
 
 **Vectorization Strategy:**
-- Only `Chunk.text`, `Chunk.keywords`, `Summary.text`, `Summary.concepts` are vectorized
+- Only `Chunk_v2.text`, `Chunk_v2.keywords`, `Summary_v2.text`, `Summary_v2.concepts` are vectorized
+- Manual vectorization with Python GPU embedder (BAAI/bge-m3, 1024-dim, RTX 4070)
 - Metadata fields use `skip_vectorization=True` for filtering performance
 - Nested objects avoid joins for efficient single-query retrieval
 - BAAI/bge-m3 model: 1024 dimensions, 8192 token context