RAG Converter - securely make your personal and work files AI readable
Convert PDF, Word, PowerPoint, Excel, Images, and Video into AI readable data for free in your browser, keep your data completely private.
Drop documents or a folder here
Everything is converted in this tab, files never leave your computer.
PDF, Word, PowerPoint, Excel, Markdown, images and audio.
Nothing is uploaded
The converter runs locally inside this tab. Your documents are read by the page, never sent anywhere.
Real formats
PDF, Word, PowerPoint and Excel are parsed directly - no LibreOffice, no cloud conversion step, no queue.
Portable output
JSONL with one object per chunk, so it loads into SpaceDB, Chroma, Qdrant, pgvector, LanceDB or anything else that takes vectors.
Learn how document conversion for RAG works
Browse all guides- What is RAG?Why documents have to become passages and vectors before an AI can answer from them.
- RAG for MarkdownWhy documentation is the best first corpus - headings are ready-made chunks.
- Local RAG in the browserConvert without uploading. What a tab can do, and where it runs out.
- What is JSONL?The portable format every vector database can import.
Local RAG, in your browser
Simple and fast for a quick conversion. Drop a file or a folder in and get chunks back in seconds - the converter is WebAssembly running in this tab, so nothing is uploaded and there is nothing to install.
- Documents, images and audio
- Token-sized overlapping chunks
- chunks.jsonl + manifest.json
API for developers
One HTTP endpoint: POST a document, get the same JSONL back. Built for pipelines and batch jobs, with the full-size models a browser cannot run - Carmenta at native resolution, Mercury for speech, Argus for video.
- POST /v1/convert, multipart in, NDJSON out
- No file limit
- Identical output structure to the free tier
Questions
What does RAG Converter do?
RAG Converter turns documents into embeddings that an AI can search. It reads PDF, Word, PowerPoint, Excel, Markdown, image and audio files, splits them into overlapping passages, and exports JSONL you can load into any vector database.
What is RAG, in plain terms?
RAG stands for retrieval-augmented generation. It makes your personal and work files AI readable: the files are split into passages and turned into vectors, so an AI assistant can look up the relevant passage and answer from your documents instead of guessing.
Are my documents uploaded anywhere?
No. The free converter is compiled to WebAssembly and runs inside your browser tab, so your documents are read by the page and never sent to a server. The page does ask our server how many free conversions remain, and that request carries only a number.
Is there an API for developers?
Yes. POST a document to /v1/convert and the same JSONL comes back, so a script and a person get an identical artifact. The API also runs the full-size models a browser cannot - OCR at native resolution, speech transcription and video - and has no file limit. It needs a sign-in; the browser converter does not.
Can I convert a PDF to Markdown for an LLM?
That is the same job, one step earlier. Markdown is usually an intermediate format on the way to chunks, so RAG Converter goes straight from your PDF to sized, overlapping chunks with a manifest - there is no intermediate file to get the chunking wrong on. PDFs with a text layer convert cleanly; scanned PDFs need OCR.
What chunk size should I use for RAG?
Size chunks to the embedding model that will encode them, not to the language model's context window - anything past the encoder's input window is silently discarded rather than embedded. 256 to 512 tokens with 10 to 15 percent overlap is a reasonable default for sentence-transformer models.
Does it do OCR on scanned documents and images?
Yes, in the browser. Carmenta is compiled to WebAssembly and reads images in your tab, so a scan or screenshot becomes text without the file leaving the page. The API runs the same engine at full resolution for larger jobs.
What file formats are supported?
PDF, Word (.docx), PowerPoint (.pptx), Excel (.xlsx), Markdown, plain text, HTML, XML, CSV, JSON, images (JPEG, PNG, GIF, BMP, WebP) and audio (WAV, FLAC).
What is JSONL and why use it for RAG?
JSONL is one JSON object per line. There is no formal interchange format for RAG corpora - every vector database has its own binary format - but nearly all of them import newline-delimited JSON, which makes it the practical standard for moving chunks and embeddings between tools.
Which vector databases can I load the output into?
SpaceDB, Chroma, Qdrant, pgvector, LanceDB, Weaviate, Pinecone and Milvus all accept the exported JSONL, either directly or through a short import script. Each line carries an id, the chunk text, its embedding and metadata.
How large can a chunk be, and do chunks overlap?
Chunks are about 256 tokens with roughly 32 tokens of overlap, cut on paragraph and sentence boundaries rather than mid-word. The overlap means a fact spanning a boundary still appears whole in at least one chunk.
Is RAG Converter free?
Yes. The browser converter is free with no account, because the work happens on your own machine. A paid API is available for larger corpora, semantic embeddings, an entity graph and a stored corpus you can query.