Zvec Vector Database: In-Process Embedded Search in C++

Milvus is a cloud-native vector database built for scalable Approximate Nearest Neighbor (ANN) search, implemented in Go and C++ to use hardware acceleration for CPU/GPU. Its fully-distributed and Kubernetes-native architecture separates compute and storage, enabling horizontal scaling to manage billions of vectors and tens of thousands of search queries with real-time updates. Milvus supports various vector index types, including HNSW and DiskANN, and features like metadata filtering, range search, and both dense and sparse vectors for hybrid search with full-text capabilities (e.g., BM25). Developers can deploy it in a standalone mode, use Milvus Lite for Python quickstarts, or opt for the managed Zilliz Cloud service. The database also provides multi-tenancy, hot/cold storage, and data security through user authentication, TLS encryption, and Role-Based Access Control. It integrates with AI tools such as LangChain and LlamaIndex, and offers ecosystem components like Attu for GUI administration and connectors for Spark, Kafka, Fivetran, and Airbyte, targeting applications like Retrieval-Augmented Generation (RAG) and recommendation systems.

Typesense is a fast, typo-tolerant in-memory search engine written in C++ that positions itself as a more developer-friendly alternative to Elasticsearch and a self-hostable replacement for Algolia. The core appeal is that it ships as a single binary with no runtime dependencies, making it straightforward to deploy locally or in production with a single command, and it handles fuzzy matching out-of-the-box without configuration. Beyond standard full-text search, it includes vector and hybrid search capabilities, geo search, faceting, filtering, synonyms, and a built-in RAG feature for conversational search over your own data. The API is clean and well-documented, with official clients for JavaScript, PHP, Python, and Ruby, plus a large ecosystem of community clients and framework integrations covering Laravel, Django, Ruby on Rails, Firebase, WordPress, and more. Benchmarks against real-world datasets show solid throughput (up to 250 concurrent queries per second on a 3-node cluster with 3 million products) and low latency (under 30ms average at scale), though the memory footprint scales with your data — plan for roughly 165MB per million records depending on field types and schema. The trade-off is that it lacks some enterprise-grade analytics features present in Algolia, and the GPL-3.0 license means server software usage is unrestricted, but modifications distributed to others must be open source. If you need search that just works without becoming a search infrastructure expert, Typesense is worth the install.

The llm-app repository from Pathway provides ready-to-deploy LLM application templates that handle the full RAG stack without requiring separate infrastructure like a standalone vector database, cache, or API framework. Built-in indexing uses usearch for vector search and Tantivy for full-text search, both operating in-memory with caching. The templates connect natively to Sharepoint, Google Drive, S3, Kafka, PostgreSQL, and real-time data APIs, keeping the vector index synchronized as source data changes. Several specialized pipelines are included: a basic question-answering RAG, live document indexing for use as a retriever backend with Langchain or LlamaIndex, multimodal RAG using GPT-4o for parsing PDFs with charts and tables, an unstructured-to-SQL pipeline for querying financial documents via natural language, Adaptive RAG which the README claims reduces token costs up to 4x while maintaining accuracy, a fully private deployment using Mistral and Ollama, slides indexing, and video RAG via TwelveLabs. All apps expose an HTTP API and can run as Docker containers, with optional Streamlit UIs for quick testing. The MIT-licensed codebase is written primarily as Jupyter Notebooks, and while it scales to millions of document pages per the README, the Rust engine bundled into the Pathway Python library handles the heavy lifting underneath.

LlamaIndex is an open-source Python framework designed to help developers build LLM applications by augmenting large language models with their private data. Described as a "data framework," it provides tools for ingesting existing data sources and formats like APIs, PDFs, and SQL. It then helps structure this data into indices and graphs, offering an advanced retrieval and query interface to generate knowledge-augmented output. Developers can start with the `llama-index` package, which includes core components and selected integrations, or use `llama-index-core` and add specific integration packages from LlamaHub, where over 300 are available for various LLM, embedding, and vector store providers. The framework also integrates with other application frameworks such as LangChain, Flask, and Docker. Separately, LlamaParse is an associated enterprise platform focused on document agents and agentic OCR, providing services like parsing, extraction, and indexing, which can be used alongside the LlamaIndex framework or on its own.