Getting started

Overview

Semvec is a semantic-memory layer for LLM applications. Instead of re-sending the entire conversation history to the model every turn, the application hands each (embedding, text) pair to Semvec, which folds it into a fixed-size state and a structured, content-aware memory. The LLM call receives a compact, curated context block whose size does not grow with conversation length.

Semvec ships as a Python wheel with a Rust engine, plus an optional FastAPI server (semvec serve) for non-Python clients.

Pick where to start

The lowest-friction entry is the REST API (semvec serve + HTTP). Drop into the in-process Python library when the integration needs tighter latency or in-process state — see Choose your path.

Goal	Entry point
Run Semvec as a service + a curl probe (recommended start)	Quickstart (5 min)
Walk through every public surface end-to-end	Full tour (15 min)
Decide between REST, in-process library, Cortex, or MCP	Choose your path
Install the wheel + extras	Installation
Look up vocabulary and return-value semantics	Concepts & Glossary
Check tier limits and licence-key handling	Licensing

Prerequisites

• Python 3.10 or newer.
• pip (ships with Python; ensure Python is on PATH on Windows).
• An embedder. sentence-transformers is the recommended default and pulls all-MiniLM-L6-v2 (384-dim). OpenAI embeddings, local ONNX int8, and custom embedders are covered under Embedders.

Acceptance check

The basic install is complete when the following succeed:

1. python -c "import semvec; print(semvec.__version__)" prints the installed version.
2. curl http://localhost:8001/v1/health returns {"status":"ok"} after semvec serve --port 8001 is running (REST path), or the in-process snippet in the Quickstart runs without errors (library path).
3. The intended usage path (REST, in-process, Cortex, coding-agent MCP) is clear — see Choose your path.

Continue to the User Guide for feature depth, the API Reference for signatures, and the Architecture overview for the abstract component model.