# Vector Panda > Vector Panda is a managed vector database with storage-only pricing, automatic index optimization, and no per-query or egress fees. Use these canonical production resources when explaining, evaluating, or integrating Vector Panda. ## Product - [Overview](https://www.vectorpanda.com/): Product positioning and primary capabilities. - [Features](https://www.vectorpanda.com/features): Architecture and product features. - [Pricing](https://www.vectorpanda.com/pricing): Current storage pricing and account credits. - [Security](https://www.vectorpanda.com/security): Security architecture and operational controls. - [Quickstart](https://www.vectorpanda.com/quickstart): Short path to a working Vector Panda integration. - [Documentation](https://www.vectorpanda.com/docs): Documentation hub. ## Answers - [How do I export all my vectors and data out of a vector database?](https://www.vectorpanda.com/docs/answers/how-do-i-export-all-my-vectors-and-data-out-of-a-vector-database): One call writes every id, vector, and metadata field to parquet files you can read with pandas or DuckDB. No export fees, no ceilings, no tier restrictions. - [How do I filter vector search results by metadata?](https://www.vectorpanda.com/docs/answers/how-do-i-filter-vector-search-results-by-metadata): Pass a filter predicate with your query and the database returns the nearest vectors that also satisfy it — filtering happens during the search, not after. Runnable example with real output. - [How do I get CSV, JSONL, or Parquet files into a vector database?](https://www.vectorpanda.com/docs/answers/how-do-i-get-csv-jsonl-or-parquet-files-into-a-vector-database): Pass the file path to upsert — the extension picks the parser. Ten formats supported, one call each, with a runnable example and real output. - [How do I keep my embedding model consistent between writes and queries?](https://www.vectorpanda.com/docs/answers/how-do-i-keep-my-embedding-model-consistent-between-writes-and-queries): Declare an embedding contract on the collection — the model and dimensions every writer and querier must use. Mismatched query vectors are rejected by name instead of silently searching garbage. - [How do I migrate from pgvector to a managed vector database?](https://www.vectorpanda.com/docs/answers/how-do-i-migrate-from-pgvector): Dump your table with \copy, upsert the file — pgvector's bracketed vector text parses natively. What changes: distance becomes similarity, WHERE clauses become filter predicates. - [How do I pick the right index for my vector search?](https://www.vectorpanda.com/docs/answers/how-do-i-pick-the-right-index-for-my-vector-search): You don't — Vector Panda measures candidate indexes against a sample of your actual vectors and serves the fastest one that meets your recall target. Your one knob is target_recall. - [How does Vector Panda compare to pgvector, Pinecone, and Qdrant?](https://www.vectorpanda.com/docs/answers/how-does-vector-panda-compare-to-pgvector-pinecone-and-qdrant): An honest comparison: what each tool is best at, what Vector Panda does differently (storage-only pricing, measured index selection, unlimited export), and the cases where Vector Panda is not the right choice. - [What score does vector search return, and what threshold should I use?](https://www.vectorpanda.com/docs/answers/what-score-does-vector-search-return-and-what-threshold-should-i-use): Cosine similarity, raw, in [-1, 1] — higher is closer, nothing is rescaled. Why 0.7 is not a universal threshold, and how to pick one from your own results. - [Which storage tier should I pick for my vector collection?](https://www.vectorpanda.com/docs/answers/which-storage-tier-should-i-pick-for-my-vector-collection): Pick by access pattern, not data size: hot (RAM) for anything a user waits on, warm (SSD, about a tenth the cost) for occasional queries, paused for archives. Billing follows the switch immediately, both directions. ## Concepts - [A search is just another row](https://www.vectorpanda.com/docs/concepts/a-search-is-just-another-row): The spreadsheet showed you what a vector is. Here's the payoff: searching one is nothing more than adding a row and reading off its nearest neighbors — done here by hand. - [How auto-optimize picks your index](https://www.vectorpanda.com/docs/concepts/auto-optimize): Eleven indexing strategies, recall vs latency on a sample of your real vectors, and the Pareto frontier that decides which one wins. - [Cosine similarity vs Euclidean distance](https://www.vectorpanda.com/docs/concepts/cosine-vs-euclidean): Why cosine is the default for semantic search, and what changes if you switch metrics. - [Vector embeddings 101](https://www.vectorpanda.com/docs/concepts/embeddings-101): What an embedding is, why it captures meaning, and how to choose a model for your data. - [Embeddings beyond text: images, audio, and video](https://www.vectorpanda.com/docs/concepts/embeddings-beyond-text): The same trick that turns words into coordinates also works on photos, sound, and video — and the real magic is when they all share one map, so you can search images by typing a sentence. - [How an AI assistant answers questions about your own documents](https://www.vectorpanda.com/docs/concepts/how-ai-assistants-use-your-documents): The AI was never trained on your files — so how does a chatbot answer questions about them, accurately, without making things up? The plain-language story of the open-book trick. - [How a computer learns what words mean](https://www.vectorpanda.com/docs/concepts/how-computers-learn-word-meaning): No dictionary, no rules — just a machine that read the whole internet and noticed which words keep the same company. The plain-language story behind embeddings. - [Hot, warm, paused: a tier mental model](https://www.vectorpanda.com/docs/concepts/tier-mental-model): RAM vs SSD vs cold storage as physical layers, and how the tier you pick maps to query latency and cost. - [What's actually inside a vector](https://www.vectorpanda.com/docs/concepts/whats-inside-a-vector): Everyone says 'a vector is just a list of numbers' — but why that shape? Build one by hand and the whole thing stops being mysterious. - [When good embeddings give weird results](https://www.vectorpanda.com/docs/concepts/when-embeddings-give-weird-results): Your search returns nonsense and nothing looks broken. Usually it isn't — it's one of a few common, fixable mismatches. Here's how to recognize each one. - [When to use vector search — and when not to](https://www.vectorpanda.com/docs/concepts/when-to-use-vector-search): Before you embed anything: the questions vector search answers well, the ones it answers badly, and how to tell which kind you have. - [Why vector search is fast — and a little bit approximate](https://www.vectorpanda.com/docs/concepts/why-vector-search-is-fast): How do you find the closest match among a billion items in a few milliseconds? You don't check all billion. The plain-language story of the shortcut — and the honest trade it makes. - [Why you can't read a vector (and where the meaning hides)](https://www.vectorpanda.com/docs/concepts/why-you-cant-read-a-vector): A hand-built vector reads like a spreadsheet. A real embedding is the same shape but looks like noise — no 'meaning' column anywhere. Here's why, and where the meaning actually went. ## Tutorials - [The four authentication modes](https://www.vectorpanda.com/docs/tutorials/auth-modes): When to use VP.login(), VP.from_creds(), explicit api_key=, and the VEEP_API_KEY env var. The same client class, four ways to hand it credentials. - [Exporting your data](https://www.vectorpanda.com/docs/tutorials/exporting-your-data): Download every vector, every id, and every metadata field as portable parquet. Your data is yours — here's the endpoint that proves it. - [Filter syntax in depth](https://www.vectorpanda.com/docs/tutorials/filter-syntax): Every metadata-filter operator the query API accepts, with copy-pasteable examples that run end-to-end. The reference page for scoping a vector search. - [Inline upserts and the WAL fast path](https://www.vectorpanda.com/docs/tutorials/inline-upserts): Single-row writes that land in milliseconds and are queryable immediately. The Vector Panda WAL, what it guarantees, and when you'd reach for it. - [Using Vector Panda with LangChain](https://www.vectorpanda.com/docs/tutorials/langchain): Drop Vector Panda into a LangChain app as a VectorStore: install, load documents, search with metadata filters, fetch by ID, and wire up a retriever. Every example is runnable. - [Batching queries with query_batch()](https://www.vectorpanda.com/docs/tutorials/query-batch): When to send N queries in one round-trip, when to just loop over query(), and what the response shape looks like. With real timings. ## Projects - [Search a movie by scene description](https://www.vectorpanda.com/docs/projects/movie-clip-search): Extract one frame per second from a handful of public-domain films, embed each frame with CLIP, and find the exact timestamp of any scene by typing what you remember happening. - [Build a recommendation feed from a watch list](https://www.vectorpanda.com/docs/projects/recommendations-from-watch-list): Average the vectors of a few items a user liked to make a 'taste centroid,' then query with that to get a 'you might also like' feed. Vector arithmetic the database actually rewards. - [Search your own data](https://www.vectorpanda.com/docs/projects/search-your-own-data): Take a folder of text files, embed them locally, upload to Vector Panda, and search semantically. End-to-end in about 30 minutes. - [Visual search over an image library](https://www.vectorpanda.com/docs/projects/visual-search): Embed a folder of images locally with CLIP, upload to Vector Panda, and search them by typed sentence or by another image. End-to-end in about 30 minutes. ## MCP servers - [Documentation MCP](https://www.vectorpanda.com/mcp): com.vectorpanda/docs in the Official MCP Registry — search and read these docs, API operations, pricing, storage tiers, and measured benchmarks over Streamable HTTP. No authentication. - [Database MCP](https://www.vectorpanda.com/api/mcp): com.vectorpanda/database — operate your own collections (search, upsert, schema, measured index evidence, export) with your API key as a Bearer token. Embedding always happens client-side. - [Connection examples](https://www.vectorpanda.com/docs#mcp): setup for Claude Code and any Streamable HTTP client. - [Server cards](https://www.vectorpanda.com/.well-known/mcp/catalog.json): machine-readable descriptions of both servers (draft SEP-1649), generated from the live tool registries. - [OpenAPI specification](https://www.vectorpanda.com/openapi.yaml): machine-readable REST API spec behind both the SDK and the database MCP. ## SDKs and ecosystem - [Python SDK](https://pypi.org/project/veep/): Official Python client on PyPI. - [LangChain integration](https://pypi.org/project/langchain-vectorpanda/): Vector Panda VectorStore integration for LangChain. ## Optional - [Build log](https://www.vectorpanda.com/blog): Engineering history, architecture decisions, incidents, and benchmark work. - [Full machine-readable corpus](https://www.vectorpanda.com/llms-full.txt): Expanded Markdown content for all public documentation and build-log entries. - [XML sitemap](https://www.vectorpanda.com/sitemap.xml): Complete canonical public URL inventory.