bench get-app https://github.com/amsys/crema

[![Listed on Frappe Gems](https://frappegems.com/api/method/frappe_gems.seo.badge?app=amsys%2Fcrema)](https://frappegems.com/gems/apps/amsys/crema)

Frappe Crema

A hardened LLM interface and security layer for Frappe apps — not a chatbot.

Alpha. Crema is version 0.0.1. The public API, the doctype fields, and the interface set can change without a migration path. Do not run it in production yet.

Frappe Crema gives every app in your bench one route to OpenAI-compatible LLM providers — OpenAI, OpenRouter, Groq, Mistral, DeepSeek, Ollama, and more. Named interfaces (one per use-case: translation, OCR, extraction, ...) bind a provider, a model, a system prompt, an isolation user, and a security layer. There is no other way to reach a provider from application code — every call goes through from crema import ask.

Crema is a security and integration layer, not a conversational agent. Building a chatbot on top is explicitly out of scope — see ROADMAP.md.

Quick start

1. Add a provider. Open Crema Settings (Crema's default landing page), click New from Template in the Providers panel, pick a preset, paste your API key, and create — this wizard also sets it as the default provider.
2. Set the default model. Pick a Default Model above the Model Assignments grid. A Default Isolation User is already filled in (install creates one for you). Save — every interface now works; the grid stays collapsed unless you open a row to override one interface's provider, model, or isolation user individually.

3. Call it:

   from crema import ask
   ask("Fix grammar: helo wrld")
   

See docs/ for full setup, usage, and automation guides.

Desk UI

A robot button on every list view and every form, plus Ask … in the awesomebar. Runs as the signed-in user, never a service account.

• List → view. Type a request to filter, sort, limit, group, or switch to Report or Kanban. If the result is empty, one more database query finds which text field holds your words and filters on that — no second model call.
• List → records. Drop a document to get one or more prefilled new records (the model decides how many).
• Form → diff. Type an instruction to get a proposed diff for the open document.

Nothing is written until you save it. A blocked prompt, a budget cap, or any other failure surfaces as a message — never a silent no-op.

Interfaces

Interface	Purpose	Fallback
simple	General-purpose, lightweight calls	none — end of the fallback chain
translation	Faithful translation, preserving tone and meaning	simple
complex	Careful, thorough reasoning for demanding tasks	simple
ocr	Extract text from images / scanned PDFs	none
advanced_ocr	Stronger retry when ocr is low-confidence	none — escalation target only
security	Layer 2 guard: classifies a prompt's risk	none — fails loud if unresolved
extraction	Pull structured data out of content	complex
classification	Classify / label content	simple
summarization	Concise, accurate summaries	simple
transform	Propose a diff for an ERP document (never auto-writes)	complex
view	Turn a prompt into a List/Report/Kanban view for the desk UI	complex
transcribe	Speech-to-text for audio files	none

Features

• One entry point — no public function accepts a model, provider, or API key; only an interface name. ask() also takes context, files (File URLs or in-memory (bytes, mime) tuples), and multi-turn history.
• 12 named interfaces with an automatic fallback chain. Other apps register their own interfaces through a crema_interfaces hook; a predefined name can never be overridden.
• Three security layers — a local regex/unicode prompt scan, an optional LLM guard that classifies risk, and an output-trap nonce that catches a hijacked response.
• Isolation user sandbox — document access runs under a fenced, low-privilege Frappe user; Frappe's own permission engine enforces it, not Crema's.
• OCR for images, scanned PDFs, and text PDFs, with a confidence score and automatic escalation to a stronger model.
• Transcription — transcribe() turns an audio file into text, with the same budget checks and audit log as every other call.
• Propose, never write — extract() proposes one or more new documents from a file (how many is the model's own call); transform() proposes a diff for an existing document. Neither writes anything; the caller applies the result.
• Scheduled automation — fetch a URL, self-plan once, extract, upsert, on a cron schedule. The plan is data, never code.
• Health probe — health() and is_configured() let a consuming app show its own status page; the caller applies its own role check.
• Cost control and audit — an optional monthly USD budget per interface and per provider, response caching per interface or per call, a usage dashboard, and a full audit log: interface, model, user, status, duration, tokens, cost, and a prompt hash — never the prompt or document content itself.
• HTTP endpoints for ask, extract, and transform, rate-limited per IP and per user.
• Admin in one place — Crema Settings holds the providers (a template wizard for OpenAI, OpenRouter, Groq, Mistral, DeepSeek, or Ollama, with live model autocomplete and connection checks) and the per-interface model assignments; a default provider, model, and isolation user cover every interface out of the box. A Desk workspace adds the doctypes and the usage report.

Requirements

Requirement	Notes
Frappe v16	installed and managed by bench
Python 3.14
litellm	the provider call layer
pymupdf	PDF text extraction and rendering
croniter	validates automation schedules — ships with Frappe itself

Documentation

Full guides live in docs/: install, configure, use, automate, and the security model.

License

MIT — see LICENSE.