Every retrieval-augmented generation system needs an ingestion layer — something that takes content from the world and transforms it into a form suitable for embedding, chunking, and retrieval. For web content, this ingestion problem is surprisingly hard. A typical web page is 70–90% noise: navigation menus, advertisement containers, cookie consent banners, JavaScript bundles, social share widgets, footer links, and tracking pixels. The actual article, documentation, or product content that you want to index is buried in this structure, and simple HTML stripping produces text that is nearly as noisy as the original HTML.
Reader-LM solves this problem with a model specifically trained to distinguish meaningful content from structural and decorative noise. Given raw HTML, it returns clean markdown — headings preserved, code blocks formatted, tables structured, lists maintained, and boilerplate removed. The result feeds directly into the embedding step of a RAG pipeline without requiring the additional preprocessing steps that general-purpose HTML cleaning tools typically need.
Reader-LM 2 is the current version, released by Jina AI in 2024. It is available in 0.5B and 1.5B parameter versions, both small enough to run on CPU-only hardware at practical speeds for batch and real-time ingestion workflows. The model supports documents up to 256K tokens, handling even the longest technical documentation pages and dense research articles. Private deployment means the URLs you are ingesting, the content of those pages, and the structure of your knowledge base never leave your infrastructure.
Why the Ingestion Layer Matters as Much as the LLM
RAG system quality is determined at every stage of the pipeline, and the ingestion and extraction stage is where most production systems underperform. An embedding model working on poorly extracted, noise-contaminated text will produce embeddings that don't reliably cluster by meaning. A retrieval system working on those embeddings will return chunks that contain the right topic surrounded by irrelevant navigational content. And the LLM generating the final answer will either ignore the noise and miss context, or include it and produce confused, poorly-grounded responses.
Reader-LM addresses this at the root. Clean extraction means cleaner chunks, better embeddings, more reliable retrieval, and ultimately more accurate generated answers. For organizations investing in private RAG infrastructure, the extraction layer is often the most neglected component and the one that produces the most improvement when done properly.
The privacy dimension of the ingestion layer is also underappreciated. Organizations building competitive intelligence systems, research monitoring tools, or internal knowledge bases often use non-public or sensitive URLs as sources. Routing these through Jina AI's public reader API — or any external service — means revealing what you are monitoring, researching, or indexing to a third party. Private Reader-LM deployment means your research targets and ingest patterns stay internal.
Where Reader-LM Consistently Delivers
RAG pipeline web ingestion
As the first stage in a retrieval pipeline, Reader-LM converts crawled or fetched HTML into clean markdown that feeds directly into chunking and embedding steps. The structured output — with preserved headings, code blocks, and tables — produces chunks that embed more meaningfully than raw-text extraction.
Technical documentation indexing
API reference pages, developer documentation, and technical wikis are structurally complex HTML — nested navigation, code samples, version selectors, and cross-reference links. Reader-LM extracts the actual content while preserving the code block structure and heading hierarchy that makes technical documentation useful when retrieved.
News and research article extraction
News pages, academic preprint servers, and research publication sites wrap thin content in heavy advertising and tracking infrastructure. Reader-LM reliably identifies and extracts article content, bylines, dates, and structured body text while discarding the surrounding noise.
E-commerce and product page extraction
Product descriptions, specifications, pricing information, and structured attributes from e-commerce pages — extracted cleanly for competitive monitoring, catalog building, and price tracking systems. The consistent markdown output simplifies the downstream parsing and database population steps.
Internal wiki and intranet content indexing
Confluence pages, Notion exports, SharePoint documents, and internal wikis are often HTML-heavy and structurally inconsistent. Reader-LM normalizes this content into clean markdown suitable for a private enterprise search and RAG system — without routing internal documentation through any external service.
Competitive intelligence monitoring
Monitoring competitor websites, press release pages, regulatory filings, and job postings requires ingesting content that reveals what you are tracking. Private Reader-LM deployment means your monitoring targets stay confidential while the extraction quality supports reliable change detection and content analysis.
Who this model family is for
Who This Model Family is For
Reader LM is a strong choice if you want to build a clean, private reading archive from web articles, documentation, or saved pages. It converts messy HTML into well-structured Markdown by removing ads, navigation, sidebars, and other boilerplate. This makes it ideal for personal knowledge bases (Obsidian, Logseq, Markdown files) where you want readable, searchable content without the noise. Keep in mind it is a specialized conversion tool — it does not summarize, answer questions, or act as a chatbot.
Teams that maintain internal wikis, documentation portals, or content libraries will find Reader LM useful for standardizing messy web content into consistent Markdown suitable for ingestion into knowledge platforms or RAG systems. It is particularly helpful when migrating legacy HTML documentation or scraped web content into modern systems. Running it through NoCloudGPT keeps all sensitive internal or customer-facing pages on your own infrastructure during conversion.
Developers building content pipelines, RAG preprocessing steps, or documentation tooling will appreciate that Reader LM is purpose-built for the HTML → Markdown conversion task. It handles boilerplate removal, preserves headings/links/structure, and produces output that downstream embedding models can use effectively. It is best paired with crawlers or ingestion jobs rather than used as a general reasoning or chat model.
Researchers building corpora from web sources benefit from Reader LM's ability to produce clean, consistently formatted text that is much easier to chunk, embed, and analyze than raw HTML. It excels at extracting meaningful body content from noisy pages, which improves dataset quality for retrieval or fine-tuning workflows. It should be treated as a focused extraction component in a data pipeline, not as an analysis or labeling model.
Licensing information
Licensing Information
Reader LM is released by Jina AI. Before deploying it commercially, you should review the exact license terms directly on the official model card, as details regarding commercial use and redistribution are maintained by the model authors.
Official model card:
Disclaimer: NoCloudGPT provides this licensing summary as a convenience reference only. It does not constitute legal advice. License terms can change. Always review the official model card and consult your legal counsel before deploying any model in a commercial or regulated context.
Available model sizes
Available Model Sizes
| Model Size | Ollama Tag | Recommended NoCloudGPT Tier | Notes | |
|---|---|---|---|---|
| Reader LM 0.5B | reader-lm:0.5b |
Starter | Lightweight option for high-volume conversion where speed and low resource use matter most. Good for personal archives and simpler pages. | Get quote → |
| Reader LM 1.5B | reader-lm:1.5b |
Standard | Better at preserving structure on messy, deeply nested pages. Recommended for most production pipelines and RAG ingestion. | Get quote → |
Positioning note: Reader LM is a specialized HTML-to-Markdown converter, not a general chat or reasoning model. It performs best when given raw web/HTML content that needs to be cleaned into structured Markdown for knowledge bases or RAG systems.
Hardware and Deployment Reality
| Configuration | Recommended instance | Best use |
|---|---|---|
| Reader-LM 2 0.5B | 1–2 GB RAM / CPU only | High-volume batch ingestion, standard pages |
| Reader-LM 2 1.5B | 4 GB RAM / CPU or light GPU | Complex pages, technical documentation, better quality |
| Reader-LM + embedding model | 8 GB Lightsail | Complete private RAG ingestion pipeline on one instance |
Reader-LM is small enough that it can run alongside an embedding model on the same instance, making a complete private RAG ingestion pipeline — fetch, extract, embed — feasible on an 8 GB Lightsail instance. For production pipelines with higher throughput requirements, Reader-LM scales horizontally across CPU instances without GPU requirements.
Deploy Reader-LM on Your Infrastructure
NoCloudGPT configures Reader-LM as part of a complete private RAG stack or as a standalone extraction service, paired with your choice of embedding model and vector store. Pipeline integration for common RAG frameworks including LangChain, LlamaIndex, and Haystack is available.