01 Generate
Write a draft
For each topic, Forge produces a structured knowledge article using the model you choose - local when possible, cloud only when explicitly configured.
Forge - .elf format - browser platform
The system behind portable AI knowledge.
webelves is a browser-native AI stack built around one idea: knowledge tools should run locally, cite their sources, and export into files you can keep. This page holds the full picture - Forge, NotebookELF, the .elf format, storage, and the rules we will not compromise on.
Forge
A synthesis engine that builds knowledge files.
Give Forge a list of topics. It writes, grades, refines, and packages accepted work into portable .elf files anyone can open in a browser. No server dependency baked into the artifact.
02 Judge
Grade it honestly
A discriminator scores the draft against a rubric. Pass or fail, with notes. The point is structured quality control, not vibes.
03 Refine
Fix what is weak
Weak drafts receive the critique as new instructions. The loop repeats until the article passes or reaches its retry limit.
04 Package
Build the .elf
Accepted articles bundle into portable .elf artifacts with manifest, corpus, citations, runtime, and optional model behavior.
Output A
Portable .elf files
Each topic becomes a self-contained knowledge file for browser use, sharing, archiving, or offline review.
Output B
Training corpus
Accepted articles become a structured dataset - raw material for better small models and domain-specific generation.
Open by design
Run it, fork it, inspect it.
The .elf spec is public. Forge is Python-first. Quality rubrics are YAML. Medical literature, legal documents, field notes, private notebooks - the pipeline is meant to be controlled by the person running it, not hidden inside a hosted black box.
The format
What is inside a .elf artifact.
A single portable file - HTML in v0.2, ZIP planned for v1.0 - that bundles the knowledge base, manifest, citations, runtime, and optional model behavior needed to browse and chat offline.
Manifest
Auditable identity
Canonical JSON plus SHA-256 fingerprinting gives an artifact a stable identity. Same manifest, same fingerprint, forever.
Corpus
Sources and passages
Articles, passage chunks, and source metadata stay bundled. Answers can cite exact passages instead of hand-waving.
Runtime
Browser-native UI
The artifact opens as a browser app. It can use wllama/GGUF, transformers.js/ONNX, or non-model search depending on size and deployment needs.
Model
Optional, not required
Full offline artifacts can bundle model behavior. Lite artifacts can fetch a model once, or fall back to search-only operation.
Spec direction
v0.2 now, v1.0 later
Current demos are self-contained HTML. The v1 direction is an inspectable ZIP container with manifest, corpus, runtime, assets, and optional model files.
Why it matters
Files outlive platforms
A knowledge artifact should not require your account, our server, or a future subscription to remain readable.
NotebookELF
NotebookLM-style work, exported as .elf.
A browser studio for producing portable knowledge files: sources, grounded RAG chat with inline citations, notes, and export in one workspace.
Workspace
Sources - Chat - Studio
Upload or paste sources, ask questions, and collect notes without leaving the browser.
Citations
Trace every answer
Answers are designed around machine-readable citation spans that point back to the source passages.
Inputs
PDF, paste, URL
Documents are processed locally when possible using browser APIs, pdf.js, extraction workers, and local storage.
| Phase | Goal | Status |
|---|---|---|
| 1 | Notebook dashboard, workspace, source migration | scaffolded |
| 2 | Source ingest UX - file, paste, URL | partial |
| 3 | Auto-generate notebook identity | planned |
| 4 | Chat loop with structured citations | planned |
| 5 | Studio notes panel | planned |
| 6 | .elf runtime upgrade with citations and source viewer | planned |
Storage
Your browser is the database.
No backend database is required for the free local tools. Your notebooks, files, indexes, and artifacts live in browser storage APIs scoped to this origin.
OPFS
Origin Private File System
Large blobs, built artifacts, and persistent workspace data live in the browser sandbox for this site.
IDB
IndexedDB
Structured metadata, source records, indexes, and notebook state survive restarts without a server.
Local
Settings and keys
Local preferences and user-provided API settings can stay in the browser. Clear site data removes them.
Backup rule: because local data belongs to the browser, clearing site data can delete notebooks. Export ZIP backups before cleaning browser storage or switching machines.
Hard commitments
What we will not compromise on.
The product can grow, but these constraints are the spine of the system.
- Local firstFree tools should run at the consumer edge. Documents, queries, and outputs stay local unless a user explicitly chooses a remote service.
- Portable.elf artifacts should remain readable without depending on webelves accounts or future servers.
- AuditableThe manifest and fingerprint should make artifacts inspectable and content-addressable.
- CitedKnowledge chat should ground answers in passages, not free-floating prose.
- Open specThe .elf format and documentation should be public enough for other people to build against.
- PracticalSubscriptions fund hosted/pro workflows. They are not a license to read your own exported artifacts.
Work with us
Building AI that belongs to the person running it.
Frontend platform, browser storage, WASM, Pyodide, local models, Python pipelines, technical writing, and open specs. If that sounds like your kind of weird, reach out.
Or email directly: mesoscaleffects@gmail.com