Use Case Guide

Build Production RAG Applications

Give your model the right context before it answers. Store your documents' embeddings, retrieve by meaning, and ground every response — with no index to pick or tune.

RAG Architecture with Vector Panda

Simple flow, powerful results

📄

Document Processing

Chunk and embed your documents

→

🐼

Vector Storage

Store in Vector Panda

→

🔍

Semantic Search

Query relevant context

→

🤖

LLM Generation

Generate accurate responses

Retrieval in one snippet

Store context, retrieve by meaning, assemble the grounded prompt — runnable as-is

from veep import Client, samples

vp = Client.login()   # browser OAuth; or Client(api_key="veep_live_...")

docs = [
    "Vector databases store embeddings and search them by meaning.",
    "RAG retrieves relevant context before the model answers.",
    "Vector Panda bills for storage only; queries are free.",
]
vp.collections.create("rag-notes", tier="hot")
vp.vectors.upsert("rag-notes", vectors=[
    {"id": f"doc-{i}", "vector": samples.encode(d), "metadata": {"text": d}}
    for i, d in enumerate(docs)
])

question = "How does RAG improve an LLM's answers?"
hits = vp.vectors.query("rag-notes", samples.encode(question), top_k=2)
context = "\n".join(h.metadata["text"] for h in hits)

# Hand the assembled prompt to whichever LLM you use
prompt = f"Context:\n{context}\n\nQuestion: {question}"
print(prompt)

For the concepts behind it, read How AI assistants use your documents; the full SDK surface is in the documentation.

Measured performance

One collection on the standard Cohere 1M dataset, measured end to end through the public API with VectorDBBench. Real numbers, not projections — methodology and raw results here.

1,620

QPS sustained

Concurrent query throughput held for the full run

0.9652

Recall@100

Accuracy against exact brute-force ground truth

49.5 ms

p99 latency

Slowest 1% of queries at full load

$17.14

That collection / month

Storage-only pricing; the queries were free

Best Practices

Optimize your RAG implementation

📊

Chunk Strategically

Use overlapping chunks of 500-1000 tokens for optimal context retrieval. Include document structure in metadata for better ranking.

🏷️

Rich Metadata

Store source, timestamp, section headers, and document type. Use metadata filters to improve relevance and reduce noise.

🔄

Hybrid Search

Combine semantic search with keyword filters for precision. Use Vector Panda's metadata queries for exact matches.

📈

Dynamic k Selection

Adjust the number of retrieved chunks based on query complexity. Start with k=5 and increase for open-ended questions.

🚀

Async Processing

Ingest documents in batches — one upsert call takes a file, a DataFrame, or a list of vectors, and blocks until the data is queryable.

💾

Smart Tiering

Keep recent docs in hot storage, move historical data to warm. Save 80%+ on storage costs without sacrificing performance.