Read
document_extractor
PDF and Office documents to text, Markdown or JSON through Docling, so the model works on text rather than binary files.
pdf · docx · pptx · xlsx → text | markdown | json
AI that generates DOCX, PPTX and XLSX from company documents
DocsGPT agents generate DOCX, PPTX, XLSX, PDF and HTML files from company documents as versioned artifacts, produced by sandboxed Python code execution in which file bytes never pass through the model.
pptx · docx · xlsx · pdf · html · versioned · sandboxed python · v0.19
Last updated:
docsgpt — code_executor
quarterly-report.docx
artifact · append-only versions
- v1generated · 3 sources cited09:14
- v2regenerated with Q3 figures · v1 kept09:21current
Artifacts
Artifacts (v0.19) are files an agent produces (PPTX, DOCX, XLSX, PDF or HTML) stored as append-only versions and referenced by handle. The model never receives file bytes; it works on extracted text and passes handles between tools. Per-user quotas cap storage.
pptx · docx · xlsx · pdf · html · append-only versions · reference handles · per-user quotas
Artifact flowextract → reason → execute → generate → version
Text version
Source documents are read by document_extractor (Docling), which returns text, Markdown or JSON. The agent reasons over that text and writes code, which code_executor runs in the sandbox; artifact_generator then produces the PPTX, DOCX, XLSX, PDF or HTML file as a new, append-only version (v1, v2, …) that the user downloads or the next workflow node consumes. Only text and reference handles reach the LLM; file bytes move between the extractor, the sandbox and the artifact store and never pass through the model.
Sandboxed code execution
The code_executor tool runs stateful Python per conversation or workflow run, about 60 seconds per run, with in-code pip install. Two backends: a self-hosted Jupyter Kernel Gateway (default) or Daytona cloud VMs.
Secrets are scrubbed from the kernel environment and internal network egress is blocked, so a script can compute over the data it was given and nothing else.
code_executor
- Runtime
- stateful Python per conversation or run
- Time limit
- ~60 s per run
- Packages
- in-code pip install
- Backends
- self-hosted Jupyter Kernel Gateway (default) · Daytona cloud VMs
- Isolation
- secrets scrubbed from the kernel environment · internal network egress blocked
Four tools
Four tools do the work: one reads, one runs, one writes and one draws. Each is a normal agent tool, so an agent or a Workflow node can use them alone or in sequence.
document_extractor · code_executor · artifact_generator · mermaid.js
Run
code_executor
Stateful Python in the sandbox: transform extracted data, compute, chart, build the file. About 60 s per run, pip install allowed.
python · ~60 s · pip install · jupyter kernel gateway | daytona
Write
artifact_generator
Produces presentations, documents, spreadsheets and PDFs as versioned artifacts with reference handles.
pptx · docx · xlsx · pdf · html · v1 · v2 · …
Draw
Mermaid.js diagrams
Flowcharts, sequence diagrams and Gantt charts generated from a natural-language description.
flowchart · sequence · gantt
How it runs privately
Choose the self-hosted Jupyter Kernel Gateway backend and extraction, code execution and artifact storage all run inside your network. Daytona is a cloud sandbox option; it is convenient on Cloud and Managed but is not for air-gapped estates.
Four deployment postures, one code base. Compare deployment options →
Inside your network
- document_extractor (docling)
- jupyter kernel gateway sandbox
- artifact store · versions
- quotas
managed · on-premises · air-gapped
Cloud
Arc53 operates the sandbox and artifact store; either backend (Jupyter Kernel Gateway or Daytona) may be used.
Air-gapped
Self-hosted Jupyter Kernel Gateway only, inside your network. No Daytona, no external network.
Example
6+ h → under 5 min
IP application drafting: 6+ hours to under 5 minutes per application, with lawyer review as the final step.
Source: Balt Alnoor
Documentation
Next: Code nodes in workflows · approval before a tool runs · how documents get in · legal drafting · deployment options · pricing
Frequently asked questions
Can AI generate a PowerPoint from my documents?
- Yes. An agent reads the source documents with document_extractor, writes the deck in sandboxed Python with code_executor, and artifact_generator saves it as a versioned PPTX you download from the conversation. DOCX, XLSX, PDF and HTML work the same way.
Is code execution sandboxed?
- Yes. Code runs in a self-hosted Jupyter Kernel Gateway (default) or in Daytona cloud VMs, with secrets scrubbed from the kernel environment, internal network egress blocked and a limit of about 60 seconds per run.
Do my files pass through the model?
- No. The model receives extracted text and reference handles to artifacts; the file bytes move between the extractor, the sandbox and the artifact store.
Can I get previous versions of a generated file?
- Yes. Artifacts are append-only: every regeneration adds a version (v1, v2, …) and earlier versions stay available.
Can I run this air-gapped?
- Yes, with the self-hosted Jupyter Kernel Gateway backend, which runs inside your network. Daytona is a cloud option and is not for air-gapped estates. See air-gapped deployment.
See DocsGPT on your documents
A 30-minute demo on your use case, or a free on-premises review with an engineer.