LLM prep panel
Every conversion comes with a compact panel that gets the text ready for a model — entirely locally, with zero AI calls.
What it shows
Section titled “What it shows”- Token count with
tiktoken(o200k_base, baked into the image — works offline). - Tokens & cost saved by anonymization, so you can see what stripping PII buys you.
- Live per-model cost estimate — pricing and context windows pulled from OpenRouter (hundreds of models, cached) so the numbers are never stale.
- Context-window fit — at a glance, which models the document fits into.
- One-click RAG chunking — split into overlapping, token-bounded chunks
(
semchunk), downloadable as.jsonl. - Prompt-injection detector — flags text that tries to hijack a downstream LLM.
Simulate cost per model
Section titled “Simulate cost per model”The panel starts with no models selected — pick the ones you care about and Escriba estimates the cost of sending this exact document to each, using live pricing. It’s the fastest way to answer “which model should I use, and what will it cost?” before you spend a cent.
RAG chunking
Section titled “RAG chunking”One click splits the Markdown into token-bounded, overlapping chunks suitable for a
retrieval pipeline, downloadable as .jsonl. Useful when the document is larger than
your target model’s context window.
All of this runs on your server. No API keys, no external calls — just local math over the text you converted.