SIE Cookbook

Example apps built on open models served by SIE, the Superlinked Inference Engine. Each recipe is one job a team does every week, such as redacting chats or triaging phishing, built from SIE model calls plus rules in plain Python and YAML that you can read and change. The code runs on your machine and uses models on SIE Serverless, SIE Cluster or SIE Local.

Every recipe is its own folder with its own pyproject.toml and uv.lock, so you install only what that recipe needs. Copy a folder into your own project and it still runs.

Run a recipe

Run a recipe using SIE Serverless, SIE Cluster or SIE Local. The code runs on your machine either way. With SIE Serverless you pay for what each call uses and there are no GPUs to manage yourself: get an SIE API key and read the API docs. To run the models yourself, start from the SIE repo.

You need uv (it installs Python 3.12 for you) and, for SIE Serverless, an SIE API key (create one).

cp .env.example .env   # then set SIE_API_KEY; every recipe reads this root .env
cd security/screen_assistant_prompts
uv run recipe.py

Or open the repo in GitHub Codespaces with the badge above and add SIE_API_KEY as a Codespaces secret; the recipes read it from the environment.

A .env inside a recipe folder works too, if you copy one recipe out on its own. Each run writes its outputs to runs/<timestamp>/ in the recipe folder.

All the data in these recipes is synthetic or comes from public datasets named in each README.

Models, swaps and self-hosting

• MODELS.md lists the open models the recipes use, with licenses, and what happened when we swapped each one for another model.
• Every recipe README has a Try another model section. Each model is one line in that recipe's rules.yaml, so trying one that answers in the same format is a one-line change.
• SELF_HOSTING.md runs the recipes using SIE Cluster or SIE Local, with pip install, a docker run command or a Helm cluster. Set SIE_BASE_URL in .env; nothing else changes.
• CHOOSING_WHERE_MODELS_RUN.md helps you pick between SIE Serverless, SIE Cluster and SIE Local, and says what you agree to when several models share one GPU.

Workflow or agent?

Most recipes here are workflows: Python picks every step and calls a small model where a step needs one, so the same input takes the same steps, the cost per item is known up front, and a batch of thousands runs overnight with nobody watching. An agent is different: a large model picks its next step in a loop, calling tools as it goes. We use one only where a workflow can't do the job:

• Someone is at the screen, asking and replying. Each question depends on the last answer, and the person checks every answer before anything is sent.
• Nobody can write the steps down, but code can check the result. The agent works out the steps; a verifier it can't argue with, often the workflow itself, checks what it found.
• The workflow gets an input it wasn't built for. It hands that case to an agent overnight, code checks the agent's work, and the result goes to a person's queue in the morning.

If you can write the steps down and the job repeats, build a workflow. If you can do neither, a person keeps the job.

Four workflows also have an agent.py beside recipe.py for the jobs they can't do, each calling the workflow as a tool: KYC checks (the analyst's follow-up questions, and a one-off request from the regulator), counterparty news (the articles the workflow sends to review, overnight), the contract playbook (a lawyer's questions mid-negotiation) and repair invoices (a one-off look at one shop's claims). They need an ANTHROPIC_API_KEY too, and run with uv run --group agent agent.py.

Recipes

Run time is a warm live run using SIE Serverless, on the date in each recipe's README (2 to 4 Nov 2026 for most). A run that has to start a cold model can take several minutes longer. Calls counts the recipe's SIE requests (one request can carry many items).

Legal

Recipe	Models	SIE calls	Run time
Check contracts for the clauses a regulation requires	docling, LightOnOCR-3-4B, Qwen3-Reranker-4B, Qwen3.8-27B-FP8	96	90 s
Review a contract against your playbook	docling, gliner2-large-v1	141	34 s
Index and check a batch of documents you received	LightOnOCR-3-4B, docling, gliner2-large-v1, GLiNER2.5-Decide	39	23 s
Redact documents to the agreed rules and log every redaction	pii-redact	6	2 s
Check citations against the documents they cite	Qwen3-Reranker-4B, Qwen3.8-27B-FP8	22	19 s

Healthcare

Recipe	Models	SIE calls	Run time
Turn a faxed referral into a draft intake record	LightOnOCR-3-4B, gliner2-large-v1	48	68 s
Add outside visit notes to your chart, marked new, matching or different	gliner2-large-v1, Qwen3-Reranker-4B	48	16 s
Pseudonymize clinical notes and log every replacement	pii-redact	4	1 s
Search clinical guidance and put the passage that answers first	Qwen3-Reranker-4B	16	6 s
Draft a short internal paper from a scientist's outline, citing your own study reports and PubMed	Qwen3.8-27B-FP8, Qwen3-Embedding-4B, Qwen3-Reranker-4B	22	52 s

Insurance

Recipe	Models	SIE calls	Run time
Check a recorded call against the claim record that was saved	parakeet-tdt-0.6b-v3, gliner2-large-v1	10	5 s
Check an invoice against the claim before it's paid	LightOnOCR-3-4B, gliner2-large-v1	30	57 s
Match claimed items to receipts and list the proof still needed	LightOnOCR-3-4B, gliner2-large-v1	33	45 s
Create a coverage brief from the policy wording, with exact quotes	Qwen3-Reranker-4B, Qwen3.8-27B-FP8	19	90 s
Summarize a claim for the next adjuster, with a source for every line	pii-redact, Qwen3.8-27B-FP8	6	14 s

Finance

Recipe	Models	SIE calls	Run time
Map financial statements into your template and check the totals	docling, Qwen3-Reranker-4B	98	30 s
Check KYC documents against your policy and list what's missing	LightOnOCR-3-4B, gliner2-large-v1	48	90 s
Alert the right person when the news mentions one of your counterparties	NuNER_Zero, GLiNER2.5-Decide	7	2 s
Answer your standing questions on a new filing, next to last year's answers	Qwen3-Reranker-4B, Qwen3.8-27B-FP8	28	29 s
Keep a case summary current, with a source for every line and no card data	pii-redact, Qwen3.8-27B-FP8	6	8 s

E-commerce

Recipe	Models	SIE calls	Run time
Categorize products into your taxonomy and fill in the required attributes	Qwen3-Embedding-4B, Qwen3-Reranker-4B, gliner2-large-v1	118	45 s
Match products to your catalog: same product, variant or new	siglip-so400m-patch14-384, Qwen3-Embedding-4B, Qwen3.8-27B-FP8	15	16 s
Tag products from their photo and text against your tag lists	Qwen3.8-27B-FP8	18	34 s
Search products by photo, by words, or both	siglip-so400m-patch14-384	13	5 s
Rerank search results without breaking the shopper's filters	Qwen3-Reranker-4B	8	5 s

Retail

Recipe	Models	SIE calls	Run time
Check a shelf from one photo: in stock, right product, right price	grounding-dino-base, GLM-OCR, Qwen3.8-27B-FP8	43	43 s
Check a returned item against the sale before taking it back	GLM-OCR, Qwen3.8-27B-FP8	15	31 s
Route customer emails to the right team, escalating safety and legal	GLiNER2.5-Decide, gliner2-large-v1	2	2 s

Manufacturing

Recipe	Models	SIE calls	Run time
Fill in a work-order completion form from the technician's notes	GLiNER2.5-Decide, gliner2-large-v1	14	4 s
Check a delivery against its purchase order and certificates	LightOnOCR-3-4B, gliner2-large-v1	52	3 min
Answer maintenance questions from the equipment manual, with the page	tomoro-colqwen3-embed-4b:compact, Qwen3.8-27B-FP8	31	50 s
Check no part is missing in assembly photos	sam3	9	6 s

SRE

Recipe	Models	SIE calls	Run time
Match a burst of alerts to the runbooks that fix them	Qwen3-Embedding-4B, Qwen3-Reranker-4B, Qwen3.8-27B-FP8	12	18 s
Route alerts to the team that owns them	GLiNER2.5-Decide	5	2 s
Draft a postmortem from the incident channel, alerts and deploys, with a source for every claim	Qwen3.8-27B-FP8	7	40 s
Redact secrets and customer data from logs before you share them	pii-redact	3	2 s

Security

Recipe	Models	SIE calls	Run time
Triage reported phishing and pull out the indicators	GLiNER2.5-Decide, gliner2-large-v1	2	1 s
Screen prompts before they reach an AI assistant (copilot)	Qwen3Guard-Gen-4B, LightOnOCR-3-4B	15	10 s
Triage threat reports against what your company runs	NuNER_Zero	1	1 s
Find the attacker tools your detection rules don't cover	gliner-relex-large-v1.0	1	3 s

Agentic

These recipes also need an ANTHROPIC_API_KEY. Claude runs the agent loop, and its tools call models on SIE. The Shape column says which case from Workflow or agent? each one shows, and what checks the agent.

Recipe	Shape	Models	SIE calls	Run time
Work out why a customer's bill jumped and draft the reply, with a source for every line	Steps unknown, result checked: Python traces every figure and quote in the draft and works out the credit; the rep reads it before it's sent	Claude Opus 5.5 (Anthropic API), Qwen3-Embedding-4B, Qwen3-Reranker-4B, LightOnOCR-3-4B, pii-redact	58	850 s
Group every log line in Python, embed only what's new, and hand the odd cases to an agent	A workflow hands off what it wasn't built for: Python checks every quote and figure in the agent's note; the on-call engineer decides	Claude Opus 5.5 (Anthropic API), Qwen3-Embedding-4B, Qwen3-Reranker-4B, pii-redact	29	47 s
Fix a laptop that fails its security checks, with small models checking each step	Steps unknown, result checked: small models check each step, Python checks the report	Claude Opus 5.5 (Anthropic API), Qwen3-Reranker-4B, LightOnOCR-3-4B, pii-redact, GLiNER2.5-Decide	48	160 s
Send each coding task to the cheapest model that passes its tests	Steps unknown, result checked: the tests	Claude Opus 5.5 (Anthropic API), GLiNER2.5-Decide, Qwen3.8-27B-FP8	78	398 s
Screen job applicants blind, with an MCP server that masks who they are	Really a workflow (the first pass is the same every time) with Claude deciding; it's here to show masking enforced in an MCP server	Claude Opus 5.5 (Anthropic API), LightOnOCR-3-4B, pii-redact, Qwen3-Reranker-4B	30	117 s

Personal agents

Recipe	Models	SIE calls	Run time
Turn messages into calendar holds, payments, replies and to-dos	parakeet-tdt-0.6b-v3, LightOnOCR-3-4B, GLiNER2.5-Decide, gliner2-large-v1	30	11 s
Keep a fact sheet current from your notes, voice memos and scans	parakeet-tdt-0.6b-v3, LightOnOCR-3-4B, Qwen3-Embedding-4B, Qwen3-Reranker-4B, Qwen3.8-27B-FP8	38	26 s

Contributing a recipe

See CONTRIBUTING.md. In short: a new folder under its industry with recipe.py, rules.yaml, README.md, a pyproject.toml that lists only what the recipe imports, and a uv.lock. uvx pre-commit run --all-files runs the repo-wide checks; CI also runs the per-recipe checks listed in CONTRIBUTING.md. Open an issue to ask for a recipe.

Future work

• Scheduled live runs. CI checks every recipe without calling SIE; the Live workflow runs recipes using SIE Serverless only when a maintainer starts it. A weekly scheduled run of a few recipes would catch model drift early. It spends SIE credits, so it is not on yet.