Graph RAG bottlenecks go beyond expansion: how two-stage Jev ranks evidence
A look at Vector Graph RAG’s two-stage Jev reranking, what its 86.61% Recall@5 result measures, and how to test it on your corpus.
Published entries across all sections carrying the “RAG” tag, newest first by publication date on this site.
21 entries
A look at Vector Graph RAG’s two-stage Jev reranking, what its 86.61% Recall@5 result measures, and how to test it on your corpus.
Zilliz's open Graph RAG stores entities, relations and text as vectors in Milvus — no graph database — and two-stage Jev retrieval hits 86.61% Recall@5. MIT.
dataelement's enterprise agent platform unifies orchestration, RAG, model management and governance, with a DeepSeek Harness desktop client. Apache-2.0.
Zleap-AI's open Rust knowledge base: zero-LLM indexing and search, 14MB app, ~200ms per million characters, SAG2 cited retrieval. MIT.
LEANN, the MLSys 2026 best paper, stores no vectors and computes them at query time — 97% storage savings for fully local RAG. 201GB becomes 6GB. MIT.
AGPL-3.0 chatbot framework: QQ, WeChat Work, Feishu, DingTalk and Telegram in one WebUI, with LLM chat, MCP, RAG and a 1,000+ plugin market.
Tencent Cloud opened Yuanbao Search to developers: Express, Standard and Flagship cost ¥14, ¥20 and ¥26 per 1,000 calls, and Flagship is enterprise-only.
An open-access Scientific Reports paper runs four vibration sensors and three LLM/VLM agents over a 16th-century minaret, scoring six zones separately.
Tencent Hunyuan open-sourced Hy Memory 2.0: separate memory pipelines for chat, work and code, a 30B-A3B memory model, and MemHub bridging ten AI tools.
XERJ is a local Rust search engine: index code and docs once, then let agents retrieve passages by file and line. Elasticsearch API and MCP included.
DeepSeek opened in-house Search and Fetch APIs on Oct 17 at $0.75 per 1,000 successful calls (5 yuan in Chinese docs); tool-call billing starts Nov 1.
Dify puts workflow, RAG, agents and model management in one visual platform: hundreds of model providers, Docker Compose self-hosting, APIs for your business.
Google's EmbeddingGemma 2 led the October 12 Product Hunt daily board with 253 votes; six of the top ten sell agents data, memory or legwork.
Upstash's hosted docs service. Coding agents fetch current, version-specific docs and code examples by library name, which cuts stale or invented APIs.
Drop PDFs, Markdown and code into a folder and query them with hybrid search and reranking. Embeddings run locally: no API keys, no external services.
Qdrant's official MCP server: two tools, store and find, with text embedded by FastEmbed. Works with a local file or a Qdrant server, for memory or snippets.
awesome-llm-apps collects 100+ open-source AI agents, skills, and RAG templates that run with one API key — Apache-2.0, works with Claude, GPT, DeepSeek, Qwen.
Graphiti is Zep's open-source temporal knowledge graph: facts carry validity windows and auto-invalidate, so agents recall by meaning, relation and time.
Morphik Core is an open-source multimodal retrieval engine that embeds charts natively, 3,715+ GitHub stars, free under BSL 1.1 below $2,000/mo revenue.
OpenCrab, by Korea's AgentKorea, turns documents into evidence-backed ontology packs queryable via MCP from Claude or Cursor, paid tiers $10-30/month.
A self-hostable, TypeScript web scraping toolkit that covers everything from static pages to JS-rendered pages across three engines, with bulk search-result collection, full-site crawling, and LLM structured extraction, all outputting LLM-friendly data.