test: image ingestion across all supported vector databases

Summary

After discovering and fixing image ingestion issues in Milvus and Weaviate, we should systematically test image ingestion across all supported vector databases to identify and fix any remaining issues.

Current Status

Database	Status	Issues Found	Resolution
Milvus	❌ Blocked	64KB JSON field limit (#19)	Unfixable (Milvus limitation)
Weaviate	✅ Working	Metadata type mismatch (#20)	✅ Fixed (v0.8.2-71-g596e682)
Chroma	❓ Untested	Unknown	TBD
Qdrant	❓ Untested	Unknown	TBD
Supabase	❓ Untested	Unknown	TBD
MongoDB	❓ Untested	Unknown	TBD
Neo4j	❓ Untested	Unknown	TBD

Test Plan

Test Dataset

File: data/tamarkin/test/2022-tamarkin-auction-catalogue.pdf

• Location: auctionsmax-ai repo
• Size: 4.9MB
• Content: 26 text chunks + 253 images
• Image sizes: 5KB to 81KB (tests size limit handling)

Test Procedure (Per Database)

1. Setup

# Ensure database is running (Docker or cloud)
# Configure in config.yaml

2. Create Collections

weave cols create AuctionListings --text --flat-metadata --<db-name>
weave cols create AuctionImages --image --flat-metadata --<db-name>

Expected: Collections created successfully

3. Ingest PDF

weave docs create AuctionListings \
  data/tamarkin/test/2022-tamarkin-auction-catalogue.pdf \
  --image-collection AuctionImages \
  --<db-name> \
  --batch-size 10

Expected:

• ✅ 26/26 text chunks created
• ✅ 253/253 images created

4. Verify Storage

weave docs count AuctionListings --<db-name>
# Expected: 26 documents

weave docs count AuctionImages --<db-name>
# Expected: 253 documents

5. Test Search

weave cols query AuctionImages "Leica M3 camera" --<db-name>

Expected: Returns relevant image results

Issues to Look For

Based on Milvus and Weaviate findings, check for:

1. Size Limits

   • JSON field limits
   • VARCHAR/String field limits
   • Binary field limits
   • Does it handle 81KB images?

2. Schema Issues

   • Metadata format (string vs object)
   • Vector configuration
   • Index type specification
   • Field type mismatches

3. Storage Issues

   • Base64 encoding overhead
   • Image data truncation
   • Silent failures

4. Performance

   • Ingestion speed
   • Memory usage
   • Batch processing

Databases to Test

Priority 1 (Common Choices)

Chroma (local + cloud)

• Docker available: Yes
• Cloud available: Yes
• Expected issues: Unknown

Qdrant (local + cloud)

• Docker available: Yes
• Cloud available: Yes
• Expected issues: Unknown

Priority 2 (Enterprise)

Supabase (PGVector)

• Docker available: Yes
• Cloud available: Yes
• Expected issues: Possible PG JSON limits

MongoDB Atlas (Vector Search)

• Docker available: Yes
• Cloud available: Yes
• Expected issues: BSON size limits?

Neo4j (Vector + Graph)

• Docker available: Yes
• Cloud available: Yes
• Expected issues: Property size limits?

Success Criteria

For each database:

• Collections create successfully
• All 26 text chunks ingest
• All 253 images ingest (including 81KB image)
• Document counts match expected
• Search queries work
• No silent failures or truncation

Deliverables

1. Test Results Document showing status for each DB
2. Issues created for any bugs found
3. Documentation updates with DB recommendations
4. Schema fixes where needed

Timeline

• Phase 1 (Priority 1 DBs): Test Chroma + Qdrant
• Phase 2 (Priority 2 DBs): Test Supabase + MongoDB + Neo4j
• Phase 3 (Documentation): Update docs with recommendations

Recommendation Matrix (Goal)

After testing, create a recommendation matrix:

Use Case	Recommended DB	Why
Small images (<64KB)	Milvus	Fast, mature, widely used
Large images (>64KB)	Weaviate	No size limits, proven working
PG-based stack	Supabase	Native PG integration
Graph + Vectors	Neo4j	Hybrid graph+vector queries
Cloud-first	Qdrant/Chroma	Managed cloud options

---

Status: Planning phase
Next Action: Begin testing Chroma (Priority 1)

Start with data/tamarkin/test/2022-tamarkin-auction-catalogue.pdf and the documented weave cols create, weave docs create, weave docs count, and weave cols query commands. Test Chroma and Qdrant first, then the Priority 2 databases, recording collection creation, 26 text chunks, 253 images, counts, search results, size-limit behavior, and any issues found.