Convert Business Files To AI-Readable RAG Without A Cloud Parser

Everyday business files are already written. Contracts, decks, invoices, meeting recordings, a scan of a whiteboard. What they are not is AI-readable. A language model cannot search a .docx the way it searches a vector. Retrieval-augmented generation is the missing step: split the file into passages, embed those passages, load them into a store your assistant can query.

Most "make my docs AI-ready" products take the file first. The parser lives in their region. The embedding job lives on their GPU. The corpus becomes their problem and your vendor assessment. RAG Converter is the other shape. The free converter is WebAssembly in the tab: PDF, Word, Excel, slides, images, and audio become overlapping chunks and JSONL on the machine that already holds the files. Developers who need a pipeline POST the same document to /v1/convert and get the same artifact back.

This is the business-file write-up. The privacy-and-WASM deep cut is RagConverter: Private RAG Conversion. The product page on this site is RAG Converter. The engines underneath sit on Remade with Rust.

Why Everyday Business Files Still Need A RAG Converter

A shared drive is not a retrieval index. SharePoint, Drive, and a zip of last quarter's board pack are storage. RAG is a second representation: token-sized passages, overlap so a fact that straddles a heading is not split in half, embeddings sized to the encoder — not to the chat model's context window. Anything past the encoder's input is discarded, not "almost embedded."

Cloud parsers fail the business test before they fail the model test. Legal, HR, finance, and clinical notes are the files people actually want an assistant to know. Sending them to a hosted ingest API creates a processing copy you then have to justify. The Federal Trade Commission has spent years on products that collected more than they needed "to make the feature work." The EFF's AI work is the same story from the other side: the processing is the disclosure.

Trust, Security, and Incentive apply to a converter the same way they apply to a vault. Trust that the free path never sees the bytes. Security is a smaller surface: WASM in-tab, or an API you choose when the job is too large for a browser. Incentive is a file you keep — JSONL that loads into SpaceDB or anyone else's store — not a corpus locked behind one vendor's query meter.

A RAG converter is how everyday business files become something an assistant can cite. Storage was never that.

How A RAG Converter Makes Business Files AI-Readable

The job is four stages: parse, chunk, embed, serialize. Formats change; the artifact does not. RAG Converter reads the file, splits on paragraph and sentence boundaries into chunks of about 256 tokens with roughly 32 tokens of overlap, optionally runs a small embedding model, and writes JSONL — one object per line, each with an id, the chunk text, its embedding, and metadata.

PDF, Word, And Spreadsheets Become AI-Readable Passages

Office files are the daily pile. .docx, .xlsx, .pptx, PDF with a text layer, Markdown, HTML, XML, CSV, JSON. Those are not "convert to Markdown then hope the chunker respects headings." Markdown is usually an intermediate people invent because the ingest tool stopped at text. A RAG converter that goes straight to sized, overlapping chunks with a manifest does not give you a second file to get wrong.

Spreadsheets need row-aware splitting or you embed a header without the value that made it mean something. Slides need the words on the slide, not the theme XML. PDFs with a text layer convert cleanly. Scanned PDFs are a different path — OCR — and belong in the next subsection, not in a silent failure that exports empty chunks.

If you are choosing a default size: 256 to 512 tokens with 10 to 15 percent overlap is a reasonable starting point for sentence-transformer encoders. Size to the embedding model. Do not size to the chat window.

Scans, Slides, And Audio Join The Same RAG Converter

A whiteboard photo and a voice memo are business files. They just do not start as text. Carmenta is the OCR engine — compiled to WASM for the browser path, full resolution on the API. Mercury is the speech path for WAV and FLAC. Argus describes images and video frames so the embedding is not just a filename.

That is the point of one converter instead of a parser, plus an ASR SaaS, plus a vision API. One JSONL schema. One load script. The NIST AI Risk Management Framework names the data-collection surface as a high-impact risk. Running OCR and ASR in the tab, or on an API you opted into, is how that surface stays a choice instead of a default upload.

JSONL That Loads Into Any Vector Database

There is no formal interchange format for RAG corpora. Every store has a binary. Nearly all of them import newline-delimited JSON. That is why the export is JSONL: SpaceDB, Chroma, Qdrant, pgvector, LanceDB, Weaviate, Pinecone, Milvus — directly or with a short import.

A proprietary bundle that only SpaceDB could read would trade one lock-in for another. The Freedom Guide is full of local-first steps that look silly next to a corpus you cannot leave. Keep the file. Load it where the product lives.

How Developers Deploy A RAG Converter On Work Files

Two surfaces, one artifact. A person drops a folder on ragconverter.com. A script POSTs the same document to /v1/convert. If those two outputs diverge, you do not have a tool — you have a demo and a production path that disagree.

Browser WASM Versus POST /v1/convert

The browser converter is free and has no account because the work happens on the user's machine. The page asks the server how many free conversions remain; that request carries a number, not the file. Use it for a desk pile, a one-off contract, a folder you do not want on a laptop that is about to go through security.

The API is for pipelines and batch jobs. It runs the full-size models a tab cannot — Carmenta at native resolution, Mercury for speech, Argus for video — and has no file-size theater. It needs a sign-in (mID / Sovereign ID). Paid tier covers larger corpora, semantic embeddings (including all-MiniLM-L6-v2), an entity graph, and a stored corpus you can query. Pick the surface by the constraint, not by the landing page.

Remade With Rust Under The Business RAG Converter

The reason a tab can ingest Office files without LibreOffice is the same program as the rest of the catalog: rusty_xml for Word, PowerPoint, and Excel; rusty_jpeg and rusty_png for images; rusty_flac for audio; rusty_alloc in WASM; SpaceDB as the storage seam; FFAI for Carmenta, Mercury, and Argus. rusty_rag is the ingest, chunk, and retrieval crate the product is built around.

That is not a logo wall. It is why there is nothing to install and no system package to bless in IT. The thesis lives in rebuilding the foundation of the internet. The inventory is The Importance of Remade With Rust. This product is what those crates look like when a developer has a folder of work files and an assistant that still guesses.

When To Convert Business Files Locally Versus In A Pipeline

Stay in the tab when the files already live on the machine that will query them, when the corpus fits a browser, and when "never uploaded" is the requirement you will be asked about. Use the API when you have a batch, a scan that needs native-resolution OCR, speech or video, or a CI job that should emit the same JSONL a human would download.

Do not wait for a perfect chunk size. Start at 256 with overlap, load the file, ask a question you already know the answer to, then move the knob. The Eras of Digital Freedom roadmap is the long product sequence. A RAG converter for everyday business files is the near-term one: make the pile AI-readable without making the pile someone else's.

If you came from the Learn index looking for "how do I get my PDFs into a vector store," the answer is the converter — browser for the desk, /v1/convert for the pipeline, JSONL for everywhere else.