I built a document chunking + embeddings API for RAG, looking for feedback
I work on production RAG systems and kept hitting the same problem: chunking quality quietly decides retrieval quality, but most setups just split text by a fixed token count and move on.
So I built a small API that does sentence-aware, structure-aware chunking and returns embeddings in one call (PDF, DOCX, CSV, XLSX, ZIP in, chunks plus vectors out, using bge-m3). The idea was to make the boring part of a RAG pipeline reliable so the rest of the stack gets cleaner inputs.
A few things I learned building it:
- Fixed-size chunking looks fine in demos and falls apart on real documents with tables and headings.
- Returning embeddings in the same call removed a lot of glue code people usually write themselves.
- Multilingual retrieval (Turkish in my case) is where weak chunking hurts the most.
It is live with a few free runs if anyone wants to try it: chunkingservice.com
I would like feedback on the chunking strategy and what formats or options are missing for your own RAG work. What are you using for chunking right now?
Hey ahmet — "fixed-size chunking looks fine in demos and falls apart on real documents with tables and headings" is a sharp framing, and how clearly that value lands on the site + first free run is exactly the kind of thing a fresh outside pass surfaces. Offer: I'll run a structured pass on chunkingservice.com — written feedback in a fixed 7-field format, back within 48h (weekend-friendly: the clock flexes, the commitment doesn't), async, no call. Not a service, a swap — in return you run a ~20–30 min pass on a signup funnel of mine, same format, so neither of us is doing a favor. Format up front: https://nonchalant-tv-881.notion.site/UI-UX-Template-Test-Version-v1-3f1e3450b0698217a20881d090a8fee0 · this runs under betapair.dev. Want me to start with yours?
Thanks for making the format and trade explicit. I can’t commit to the reciprocal 20–30 minute pass right now, so I don’t want to promise the swap. If you independently try the free run and find a reproducible issue, the most useful report would include document type, expected versus actual chunk boundaries, and the run ID.