vex — Vector Exchange

Cross-standard vector DB migration tool. Export, import, and migrate agent memory between vector stores using the open .vmig.jsonl interchange format.

npx vex export --from vektor --db slipstream-memory.db --output memories.vmig.jsonl
npx vex import --from memories.vmig.jsonl --to pinecone --api-key $KEY --index my-index --host $HOST
npx vex import --from memories.vmig.jsonl --to qdrant --collection memories
npx vex migrate --from vektor --to qdrant --db memory.db --url http://localhost:6333 --collection memories

Why

Every vector DB has a different API, a different format, and zero interop. Moving your agent memory from VEKTOR to Pinecone, or Qdrant to Weaviate, means writing a one-off script every time.

vex fixes that with a single open format and a growing connector library. Your memory is always exportable, always portable, always yours.

Connectors

| Connector | Export | Import | Status | |-----------|--------|--------|--------| | vektor | ✅ | ✅ v0.1 | Stable | | jsonl | ✅ | ✅ | Stable | | pinecone | 🔜 Phase 2 | ✅ | Tested — 4,900 vectors | | qdrant | 🔜 Phase 2 | ✅ | Tested — 3,917 vectors, auto-create | | chroma | 🔜 | 🔜 | Phase 2 | | weaviate | 🔜 | 🔜 | Phase 2 | | pgvector | 🔜 | 🔜 | Phase 2 |

Install

npm install -g vex
# or run without installing
npx vex --help

Requirements: Node.js >= 18 (native fetch required). No extra dependencies for Pinecone or Qdrant — connectors use the built-in fetch API.

Commands

Export

# Export VEKTOR memory to portable .vmig.jsonl file
vex export --from vektor --db ./slipstream-memory.db --output memories.vmig.jsonl

# Export a specific namespace only
vex export --from vektor --db ./memory.db --namespace trading --output trading.vmig.jsonl

Import

# → Pinecone
vex import --from memories.vmig.jsonl --to pinecone \
  --api-key $PINECONE_API_KEY \
  --index my-index \
  --host https://my-index-xxxx.svc.pinecone.io

# → Qdrant (auto-creates collection if missing)
vex import --from memories.vmig.jsonl --to qdrant \
  --url https://xxxx.cloud.qdrant.io:6333 \
  --collection my-collection \
  --api-key $QDRANT_API_KEY

# → Qdrant local (no auth)
vex import --from memories.vmig.jsonl --to qdrant --collection memories

Migrate (direct — no intermediate file)

# VEKTOR → Qdrant in one command
vex migrate --from vektor --to qdrant \
  --db ./memory.db \
  --url http://localhost:6333 \
  --collection memories

.vmig.jsonl Format

One JSON object per line. UTF-8. Portable across any vector store.

{
  "id": "1234",
  "text": "Pepe trending #5 on CoinGecko (+2.0% 24h)",
  "vector": [0.021, -0.043, 0.018, "...384 or 768 floats"],
  "model": "bge-small-en-v1.5",
  "dims": 384,
  "namespace": "trading",
  "score": null,
  "metadata": {
    "tags": "crypto,trending",
    "importance": 1.0,
    "agent_id": "default"
  },
  "created_at": "2025-01-15T10:23:00.000Z",
  "source_store": "vektor",
  "vex_version": "1.0.0"
}

Key decisions:

• Metadata is flat — Pinecone compatible out of the box
• namespace is top-level — structural routing, not descriptive metadata
• text field always preserved — enables cross-model re-embedding in v0.2
• score field included — useful for search-result exports
• Sidecar .vmig.meta.json for file-level metadata (record count, SHA-256 checksum, source store)

Embedding Handling

| Scenario | Behaviour | |----------|-----------| | Same model, same dims | Vectors copied directly — no re-embedding | | Dim mismatch with target | Records skipped with warning + count in summary | | null vector | Record skipped with warning | | Different model (v0.2) | Re-embed from text field via Drift-Adapter |

Connectors auto-detect target dimension (Pinecone: queries index metadata API, Qdrant: queries collection config) and filter records accordingly. Batch retry logic (3x with backoff) built in.

Sidecar Metadata

Each export and import produces a .vmig.meta.json alongside the data file:

{
  "exported_at": "2025-01-15T10:23:00.000Z",
  "source_store": "vektor",
  "record_count": 5026,
  "checksum": "sha256:abc123...",
  "vex_version": "1.0.0"
}

After import, the sidecar is updated with imported_to and imported_at fields for full auditability.

Progress & Summary

Every import shows a live progress bar and a summary block:

[████████████████████] 100% pinecone (4900/4900)

┌─ pinecone summary ─────────────────────────
│  total records   : 4900
│  upserted        : 4900
│  skipped         : 0
│  duration        : 87.3s
└────────────────────────────────────────────

Roadmap

v0.0.1 — shipped

• VEKTOR export
• JSONL round-trip
• Format spec v1.0.0

v0.1.0 — shipped

• Pinecone import (tested: 4,900 vectors)
• Qdrant import (tested: 3,917 vectors, auto-create collection)
• SHA-256 checksum in sidecar meta
• Batch retry with backoff (3x)
• Progress bar + summary block

v0.2 — next

• Pinecone export
• Qdrant export
• ChromaDB connector (import + export)
• --namespace filter on all connectors
• --limit flag for partial exports

v0.3 — shipped

• Weaviate, pgvector connectors
• Re-embedding pipeline (different model → re-embed from text field)
• Streaming for large datasets (>100k vectors)

v0.4 (premium)

• Pre-trained Drift-Adapter weights (vec2vec translation — no re-embedding required)
• Multimodal support

Contributing

PRs welcome — especially new connectors.

Each connector is a single file in connectors/ implementing two functions:

{ extract(opts), load(records, opts) }

See connectors/qdrant.js as the reference implementation. The Vex core handles batching, dimension filtering, retry, progress, and sidecar generation.

License

Apache 2.0 — free to use, fork, and build on.

Built by VEKTOR — persistent semantic memory for AI agents.