← work

sheet 02.1 · Ctrl+Space Labs · open source

Gendox

Open-source platform for AI agents on your own documents. I build most of what happens after the upload button.

role
Developer. Architecture by Chris Sekas, implementation largely mine.
period
Sep 2023 — now
stack
Java 25Spring Boot 3Spring AISpring BatchPostgreSQL + pgvectorKeycloakNext.jsRedux Toolkit
links
app.gendox.dev ↗docs.gendox.dev ↗GitHub ↗
680+
commits in gendox-core
600+
commits in gendox-devops
8
LLM providers behind one interface
01What it is

You upload documents, Gendox splits and embeds them into pgvector, and an agent answers from that knowledge base with retrieval and tool calls. The agent drops into any website with one script tag.

It is open source and self-hostable, with a live instance at app.gendox.dev.

sources · scheduled, run-lockeduploadgit repos3 bucketapiweb scrapesplitspring batchembed8 providerspgvectorpostgres 18retrieve+ tool callschat widgetinsightsdigitizationingest · async jobsserve
fig. 1 — gendox, ingest to answer (simplified)not to scale
02Ingestion: splitting and training

Documents go through Spring Batch jobs: a splitter step that cuts them into sections, then a training step that creates the embeddings. Uploads trigger the split automatically, and long jobs report their progress back to the UI instead of blocking a request.

03Integrations

Content can also arrive on its own: from a Git repository, an S3 bucket, an external API, or a web scrape. Each integration runs on a schedule with its own permissions and a run lock, so two runs never write the same documents at once.

04Document Insights

Ask the same set of questions across a whole batch of documents. Each document × question pair is a node in a task graph; the nodes run in parallel and asynchronously, and the result is a matrix with short answers, explanations, status flags and CSV export.

Getting this right meant making task execution truly async (no blocking joins), keeping answer nodes from being treated as documents, and letting supporting documents and a global prompt feed every question.

q1q2q3q4contract_a.pdfcontract_b.pdf!invoice_07.pdfreport_q3.docxcsvdonerunning!flagged for review
fig. 3 — document insights, one node per cellnot to scale
05Document Digitization

Scanned PDFs and office files are turned into text or JSON page by page with vision-capable models, with an optional JSON schema to extract structured data in one pass. No classic OCR step.

06SeaScope

An Earth Observation workspace: the agent writes a Google Earth Engine script from a question, the script runs in a browser sandbox, and the result appears as layers on a map. I built the auth flow, multi-layer support and the area-of-interest tools.

07Around it

Keycloak login and themes, subscription plans and validation, per-organization provider keys, attachments in chat, a Python SDK, and the CI/CD that deploys the whole thing to Hetzner. I am now rebuilding the frontend from scratch.

Next: Structured extraction from historical registers →Need something like this?