Skip to content

Core concepts

LLMAIx Web organizes everything under a project. Within a project you move through a fixed sequence of steps, each producing the input for the next:

Files → Preprocessing → Documents → Schemas + Prompts → Trials → Evaluation

Project

The top-level container. A project holds its own files, documents, schemas, prompts, trials, ground truth, and evaluations. Access can be scoped per user; admins may optionally see all projects.

File

A raw upload — PDF, DOC/DOCX, PNG/JPEG, CSV/XLSX, or TXT. Files are stored in local storage or S3-compatible object storage under a UUID-based filename. Duplicate uploads are detected by SHA-256 hash. See Files.

Preprocessing

The asynchronous step that turns files into text. For PDFs and images this means OCR / text extraction (four engines to choose from); for spreadsheets it means converting rows or the whole table into text. See Preprocessing.

Document

The output of preprocessing: a piece of extracted text ready to be sent to an LLM. A single file can yield one document (e.g. a PDF) or many (e.g. one per CSV row). Documents can be grouped into document sets for running trials. See Documents.

The document viewer showing extracted text on the left and a metadata sidebar on the right

A document is the extracted text (left) plus its provenance metadata (right) — the exact input a trial sends to the LLM.

Schema & prompt

A schema is a JSON schema defining the structured output you want (nested objects, arrays, all JSON types), built with a visual tree editor. A prompt pairs a system prompt (extraction rules) with a user prompt (which receives the document text). See Schemas & prompts.

Trial

A trial runs LLM extraction over a set of documents using a chosen schema + prompt + model against any OpenAI-compatible endpoint. Each document produces one trial result (the extracted JSON). See Trials.

Ground truth & evaluation

Ground truth is a spreadsheet of known-correct values. An evaluation compares a trial's results against ground truth — using per-field comparison methods (exact, fuzzy, numeric, …) — and computes accuracy metrics. See Ground truth and Evaluation.

The evaluation dashboard listing a scored trial with its overall accuracy and per-field breakdown

An evaluation scores a trial against ground truth and reports overall and per-field accuracy.

Asynchronous processing

Preprocessing and trials run as background Celery tasks. The UI shows live progress over a WebSocket connection, and both task types support cancellation (with optional rollback of anything they produced).