Vector Panda mascot, a panda in an orange hoodie and beanie, smiling and looking right toward the page content

The database for AI apps that search by meaning.

Store the numerical "embeddings" your AI model produces, search them fast, pay only for storage. $5/mo in credits to start.

Start free → View Documentation
3.1s
Time to First Query
$1.49
Per GB / Month (Warm)
$5
Credits Each Month
9+
Formats Supported

We find the best index, proven on your data

Auto-optimization dashboard: 108 of 378 tests run (29% complete, HNSW ID 1919 up next), and the best so far is PCA index ID 6659 at 100% recall, 87% faster than baseline. Scatter plot shows recall vs p95 latency across all PCA configurations evaluated.

Real run from a customer dashboard. See your own frontier in your collection's view.

Most vector databases ask you to pick HNSW or IVF or PQ, then tune a dozen knobs.
Vector Panda runs every viable strategy against a sample of your real vectors, measures accuracy, recall, and p95 latency (the slowest 5% of queries), then runs on the winner from your data's Pareto frontier — the best mix of fast and accurate.

  • PCA
  • HNSW
  • Vamana
  • IVF-Flat
  • IVF-PQ
  • PQ
  • OPQ
  • ScaNN
  • SQ
  • LSH
  • Brute-force

Built for developers

Everything you need to ship vector search, nothing you don't

Storage-only pricing

Pay per GB of vectors at rest. Queries are free, egress is free, ingest is free. $5 in monthly usage credits, no card for the first 90 days. Most prototypes finish well inside that window.

Yours to keep

Bring your own embeddings: OpenAI, Cohere, sentence-transformers, your own. Upload Parquet, CSV, NumPy, HDF5, Arrow IPC, JSONL, Safetensors, BVECS/FVECS/IVECS. Export the full collection to Parquet any time. No lock-in.

Built for the terminal

pip install veep, VP.login() opens a browser for OAuth, six lines to your first semantic search. Filter at query time with MongoDB-style predicates over your metadata. No separate filter index, no rebuilds.

Get started in 60 seconds

Six lines from install to your first semantic search

Vector Panda mascot holding a server cube, a small decorative element next to the quickstart heading
# Install Vector Panda + bundled all-MiniLM-L6-v2 ONNX encoder (~22MB)
pip install veep[samples]
# Sign in via browser, no API key wrangling
from veep import VP, samples
vp = VP.login() # or simply VP(api_key=...) instead of OAuth
# Create a collection and upload ~5,000 popular movies (Wikipedia plots)
# upsert() blocks until the data is queryable, a few seconds on a fast network
vp.collections.create("quickstart")
# Storage: ~$0.04/mo on the fastest ("hot") tier, or ~$0.01/mo on "warm". Easily within the $5 monthly credit
vp.vectors.upsert("quickstart", dataframe=samples.dataframe())
# Encode a plot description and search by meaning. Titles come back as metadata
q = samples.encode("a hobbit destroys a magic ring")
results = vp.vectors.query("quickstart", vector=q, top_k=5)
What you'll see, top 5 hits for “a hobbit destroys a magic ring”:
1. 0.5754 The Lord of the Rings: The Fellowship of the Ring (2001)
2. 0.5713 The Lord of the Rings: The Return of the King (2003)
3. 0.5400 The Lord of the Rings: The Two Towers (2002)
4. 0.3918 The Flight of Dragons (1982)
5. 0.3741 Harry Potter and the Deathly Hallows: Part 2 (2011)
Sneakier prompts (pure plot concept, no franchise words):
0.4962 “an off-duty cop is trapped in a high-rise with terrorists at christmas” → Die Hard (1988)
0.4636 “an unlikely friendship across class lines” → The Breakfast Club (1985)
0.4460 “a kid finds a stranded alien hiding in his backyard” → E.T. the Extra-Terrestrial (1982)
0.4250 “a woman is stalked by a relentless robotic killer from the future” → The Terminator (1984)

The top 3 are the entire Lord of the Rings trilogy (Fellowship, Return of the King, Two Towers) followed by other fantasy films. The query never mentions Tolkien, Frodo, or rings of power; the system found the films from the concept alone.

How samples.encode() works: ships a quantized ONNX export of sentence-transformers/all-MiniLM-L6-v2 (Apache 2.0) and runs it locally via ONNX Runtime. No embedding-API key, no network round-trips, no vendor SDK. The 5,000 movie plots come from English Wikipedia under CC BY-SA 4.0; popularity rank comes from Wikidata sitelinks (CC0). Bring your own model + data in production. This is the quickstart.

Simple, transparent pricing

Each collection picks its own tier. Match the rate to how hot the data needs to be, compartmentalize by access pattern. Change anytime, no commitments.

Hot
$5.99
per GB / month
  • In-memory storage
  • Typically < 10ms
  • Production-grade performance
Warm
$1.49
per GB / month
  • SSD storage
  • Typically < 100ms
  • Shared in-memory pool keeps recent queries fast
Paused
$0.09
per GB / month
  • Archive storage
  • Resume in minutes
  • Pre-indexed, query-ready when you resume

$5 in monthly usage credits, free to start. No strings, no card required for your first 90 days.
Add a credit card by day 90 to continue the usage credits.

Example projects

Four ways to build on the quickstart, ordered approachable to advanced

Search your own data

Swap the bundled corpus for your own. Same six-line pattern with a Parquet file, a pandas DataFrame, or a stream of inserts. Your titles, your embeddings, your search.

Recommendations from a watch list

Average the embeddings of a few items a user liked into a "taste centroid," then query with that to get a "you might also like" feed. The trick that's vector-database-only.

Visual search over an image library

Embed images with a CLIP-style encoder, then query with a typed sentence or another image. The same collection answers both — find the visual match across thousands of artworks, products, or scenes.

Search a movie by scene description

Slice public-domain films into one-frame-per-second JPEGs, embed each with CLIP, then type the scene you remember and jump to the exact archive.org timestamp.

Ready to build?

Start with $5 in monthly credits.

Get Started →