Jina AI ReaderLM-v2 enhanced release: the conversion quality from HTML to Markdown/JSON has been improved, and the Agent retrieval base has been upgraded.

Jina AI released the enhanced version of ReaderLM-v2 in June 2026. The core is to improve the conversion quality of HTML to Markdown/JSON in Reader to a higher level and continue to expand search-related capabilities. As the Agent retrieval base with r.jina.ai and s.jina.ai as the core, this update is directly related to the reliability and fidelity of the RAG system and AI Agent in obtaining web page information.

• Conversion quality enhancement: ReaderLM-v2 improves the accuracy of HTML to Markdown/JSON conversion, reducing information loss and structural confusion.
• Search capability expansion: The Search endpoint continues to be enhanced to support the Agent’s real-time information acquisition.
• Componentized retrieval base: Reader, Embeddings, Reranker, DeepSearch and other capabilities can be called independently and freely combined.
• For LLM-friendly output: Converting web page content into LLM-friendly text is a key preprocessing layer for RAG and Agent.

Version background

Jina AI is not positioned as another vector database or search engine, but a set of open source search infrastructure layers with "componentized retrieval" as its core concept. It decomposes web page reading (Reader), semantic vectorization (Embeddings), relevance rearrangement (Reranker), multi-modal understanding and deep search reasoning (DeepSearch) into API modules that can be called independently and freely combined. The Reader API has continued to evolve since 2023-01 when it exposed r.jina.ai URL reading capabilities, and ReaderLM-v2 is a key upgrade in conversion quality in this evolution.

Highlights of this version

ReaderLM-v2 conversion enhancement

• HTML to Markdown: restore the structure of the web page more accurately, remove noise such as navigation and advertisements, and retain the semantics of the text.
• HTML to JSON: Structured extraction of page information to facilitate programmatic consumption and data storage.
• Fidelity improvement: Complex structures such as tables, code blocks, and lists are less distorted during conversion, reducing downstream parsing costs.

Search and retrieval components

• s.jina.ai Search: Agent-oriented real-time search capability, working with Reader to complete the "search-read" closed loop.
• Embeddings and Reranker: Semantic vectorization and relevance rearrangement to improve retrieval quality.
• DeepSearch: Deep search reasoning capabilities, supporting multi-step retrieval scenarios.

Meaning for developers

From an industry perspective, Jina AI's ReaderLM-v2 upgrade points to the deterministic requirements for Agent infrastructure: web page information is one of the most important external data sources for Agent, and the quality of the conversion of "web pages to structured text" directly determines the answer quality of RAG applications. For domestic developers, this means that when building document question and answer, information aggregation and agent applications, they can first connect "reliable web page reading" as the infrastructure, and then add vectorization and rearrangement.

For teams building RAG framework applications such as LlamaIndex, Reader provides a high-quality implementation of "web page loader", which can significantly reduce the workload of self-built crawling and cleaning.

Tips for getting started

• RAG Application: Use Reader endpoint to replace self-built web page parsing to verify the improvement of answer accuracy due to conversion quality.
• Agent development: Combine s.jina.ai search and Reader to build an Agent information link of "real-time search - web page reading - structured output".
• Evaluation Criteria: The quality of the base is measured in two dimensions: "accuracy of structured conversion" and "relevance of search results".

Directions worthy of attention in the future

1. Boundary of conversion quality: conversion performance of complex dynamic pages, anti-crawl pages and multi-language content.
2. Real-time performance of search capabilities: The coverage of real-time performance and regional differences affects the scope of Agent implementation.
3. Comparison with domestic retrieval components: Actual measurement comparison with domestic retrieval solutions in Chinese web pages and Chinese search scenarios.