CiteCraft is the current public-facing working name for a broader AI deliverables workbench.
The codebase stays intentionally neutral while the brand is still provisional.
Turn messy sources into cited deliverables.
把杂乱资料变成带引用、可交付的成果。
This project is building an Academia-first, not academia-only workbench for source-heavy knowledge work.
它以学术场景作为首个演示入口，但底层结构从第一天就支持扩展到 research、consulting、policy 和更广的知识工作场景。
You give it PDFs, URLs, and markdown/text folders.
It gives you deliverables you can actually use:
cited briefliterature matrixslides
The product goal is not "better chat." The goal is a deterministic, traceable path from messy inputs to professional output.
这不是另一个聊天壳，也不是泛 agent 平台；它的核心是 evidence-bound deliverables。
Real screenshots from the local demo shell and poster route:
Academiatrack: journal club and literature synthesis with cited deliverables.Research / Consultingtrack: strategy brief prep from reports, market pages, and analyst notes.Trace Panel: every important output node points back to a source chunk.Poster route: a share-friendly surface built from the same real demo data.
- It ships
deliverables, not just answers. - It treats
traceabilityas a first-class product primitive. - It keeps the core deterministic and reviewable before adding agent complexity.
- It is designed so the community can later contribute templates, adapters, and deliverable types.
This repo now ships project-scoped multi-agent ops:
- repo-local subagents in
.codex/agents/ - a repo-local skill in
.agents/skills/citecraft-multi-agent/
Use them when a task spans core pipeline, demo shell, README, examples, or poster assets and you want the same exploration -> writer -> review workflow we use on the main branch.
- Every important claim should point back to a source chunk.
- Deliverables are generated as structured models first, then rendered.
- Human-readable output follows
Canonical markdown, rendered everywhere. - The web app is a demo shell, not a long-term frontend commitment.
- The golden path is the fixed sample project in
examples/academia/demo-01/.
src/workbench/ # neutral Python core
apps/web/ # demo shell
templates/ # deliverable and render templates
examples/ # sample projects and expected outputs
tests/ # unit, integration, acceptance
docs/ # contracts and architecture notes
The repo has been tested with D:\anaconda3\python.exe on this machine because the Windows Store python alias is broken.
D:\anaconda3\python.exe -m pip install -e .[dev]
$env:PYTHONPATH='src'
D:\anaconda3\python.exe -m workbench.pipeline.run examples/academia/demo-01 --output-dir examples/academia/demo-01/expected
D:\anaconda3\python.exe apps\web\app.pyThen open http://127.0.0.1:5000.
The local demo shell now supports:
- English / Chinese UI switching
- an
academiatrack and aresearch / consultingtrack - traceable preview flows for
brief,literature_matrix, andslides
Try:
http://127.0.0.1:5000/?lang=en&project=academia-demo-01http://127.0.0.1:5000/?lang=zh&project=research-demo-01http://127.0.0.1:5000/poster?lang=en&project=academia-demo-01http://127.0.0.1:5000/poster?lang=zh&project=research-demo-01
The default demo is deterministic so the 1-minute path stays stable. To try a real model through an OpenAI-compatible endpoint:
$env:WORKBENCH_PROVIDER="openai-compatible"
$env:WORKBENCH_API_KEY="..."
$env:WORKBENCH_MODEL="gpt-4.1-mini"
$env:WORKBENCH_BASE_URL="https://api.openai.com/v1"
D:\anaconda3\python.exe -m workbench.pipeline.run examples/academia/demo-01 --provider openai-compatibleIf provider config is missing or a request fails, the demo safely falls back to deterministic seed text.
P0 includes:
- source adapters for
PDF,URL, andmarkdown/text folder - a deterministic pipeline
- a narrow provider seam for future model-backed generation
- evidence traces with clear locators
- structured contracts for
brief,literature_matrix, andslides - markdown and HTML rendering
- a stable 1-minute demo path
P1 will add:
reviewer rebuttal draft- PPTX export
- stronger citation QA
- broader cross-domain examples