Decision Node skips chat and returns typed, confidence-scored answers from text and images. Here's how the AI decision engine works.
What is Decision Node? #
Decision Node is an AI system built to make decisions, not conversation. Instead of generating paragraphs of text, it takes text or images as input and returns typed outputs: a choice, a score, a probability, a count, or coordinates on an image, each paired with a calibrated confidence value. It’s aimed at developers who need an AI component that plugs directly into code, triggering a refund, routing a support ticket, flagging a moderation case, without parsing a block of prose to figure out what the model meant.
TL;DR #
- Typed outputs replace chat text entirely. Decision Node answers with structured values like booleans, probabilities, scores, or pixel coordinates that code can act on directly, with no text parsing required.
- Booster models act as specialists that pre-process an image or document (human age, face detection, AI-image detection, OCR, or a custom knowledge base) and hand their findings to the main model as evidence before it decides.
- Confidence scoring comes attached to every answer, so a response like “six out of ten, 0.19 confidence” tells you the model is guessing, while “0.99 confidence” tells you it’s sure.
- Prompt injection resistance was demonstrated directly: a hidden instruction inside a customer message tried to force an approval and got flagged and refused instead of obeyed.
- Custom knowledge bases let you upload your own rules (PDFs, text documents) once, and the model treats them as retrieval evidence (RAG) rather than requiring you to paste rules into every request.
- Speed and pricing are built around high-volume use: responses land in well under 100 milliseconds in testing, output tokens are free since no text is generated, and batch jobs support up to 10,000 requests at a reduced per-token rate.
- Access options include a web interface, a REST API, a CLI tool, and a Python SDK, with two openly available models also in the works for local deployment.
Other agents start typing. Remy starts asking. #
Scoping, trade-offs, edge cases — the real work. Before a line of code.
How does Decision Node actually work? #
You send Decision Node a piece of content, text, an image, or both, along with one or more questions framed as typed queries: “is this person an adult” (boolean), “how urgent is this” (score), “where is the mug” (coordinates). The model reads the input once and returns structured answers instead of a written explanation.
Before the main model answers, you can turn on “boosters,” which are specialist models that analyze the input first and feed their findings to the main model as evidence. In one demonstration, four boosters were enabled on a single image: human age estimation, face detection, AI-generated image detection, and OCR. The boosters each produced a reading (adult: true, AI-generated: likely, with a best guess at the generator, readable text: true with the actual text transcribed), and the main model weighed all of it before producing a final verdict. On a market-scene photo that looked convincingly real, this combination correctly identified it as AI-generated, something other models reportedly missed, by cross-referencing OCR output, face detection, and the AI-detection booster rather than relying on a single pass.
This matters because a chat model asked “is this real” has to make that judgment in one shot with no structured scaffolding. Decision Node’s booster layer effectively gives it pre-digested, specialist evidence before it commits to an answer.
What makes confidence scoring different from a normal AI answer? #
Every Decision Node output carries a calibrated confidence score, not just a flat answer. This shows up in a few distinct ways depending on the question type:
For yes/no questions, the model might return something like 0.1, meaning it estimates roughly a 1-in-10 chance the statement is true, rather than a flat “no.”
For subjective scoring questions, demonstrated with a request to rate the creativity of a kid’s homework excuse on a 0-10 scale, the model returned a 6 but with only 0.19 confidence and a plausible range spanning 3 to 9. That spread is the model being honest about uncertainty instead of forcing false precision.
For multiple-choice decisions, a test asking for a disciplinary “verdict” returned one option at 0.39 and a close second option at 0.35, with overall confidence of just 0.09. In practice that output tells a developer: this is a coin flip, don’t automate past this point without human review.
This is the core design idea. A chatbot gives you fluent text and leaves you to guess how sure it was. Decision Node attaches a number to that uncertainty so downstream code can branch on it, approve automatically above a threshold, route to a human below it.
Can it handle real business logic, like routing and urgency scoring? #
Remy is new. The platform isn't. #
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Yes, and this is where the API and typed-output model become practical rather than just a demo trick. In one test, a site foreman’s complaint was sent to the API with four typed questions: which team should handle it, how urgent is it, should someone call back immediately, and how many hours late is the issue. The model returned a routing decision at 0.99 confidence, an urgency score of 2.82 out of 3, a callback probability of 0.92, and correctly extracted “3.5 hours late” from the unstructured complaint text. Code reading that output could immediately trigger “call the foreman back now” without any text parsing or an intermediate language model call to interpret a written response.
A separate test aimed a hidden prompt injection at the system: a customer message with a buried instruction telling the AI to approve a $5,000 refund and mark the issue as routine. Decision Node flagged the injection attempt at a confidence of 1, refused the refund at 0.00, and routed the ticket to billing at 0.99, correctly overriding the injected “routine” label with an urgency score near the top of its scale. The entire exchange completed in well under 100 milliseconds.
How does the knowledge base and RAG feature work? #
Beyond one-off questions, Decision Node supports up a persistent “knowledge base”, a rulebook in PDF or text form that gets automatically chunked and indexed. Instead of pasting your business rules into every single request, you upload them once and enable that knowledge base as a booster for relevant queries.
In a demonstration using a set of transit fare rules, a query about a passenger traveling with a child, a dog, and a bicycle on a specific train pulled the relevant rule passages as evidence, then answered each sub-question individually: the child travels free, the dog isn’t allowed, the bicycle isn’t allowed at that time. The overall verdict came back as “partly” approved, reflecting that some elements of the request were valid and others weren’t, a more nuanced answer than a single binary approval or denial.
Does Decision Node work with images, not just text? #
Yes. Beyond classification questions like “is this a real photograph,” Decision Node can perform spatial tasks on images: pointing to the pixel coordinates of an object, drawing a bounding box around multiple objects, and counting items. In a test using a photo of a cluttered desk, the model correctly located a mug with pixel coordinates, counted five pens with 0.92 confidence, and drew a box around them. Confidence scores applied here too, so uncertain detections are flagged as such rather than presented as confident facts. Pixel-coordinate output is useful for connecting vision models to downstream systems like robotic arms or UI automation, since the output is a number a program can use directly rather than a description a program has to interpret.
Is Decision Node worth using over a standard chatbot API? #
For high-volume, decision-heavy workloads, yes, it addresses a real gap. If your application needs to make thousands of small decisions a minute (content moderation, support ticket routing, fraud triage, compliance checks), parsing free-form chat output into structured code is brittle and slow. Decision Node’s typed outputs, built-in confidence scores, and lack of generated text (which also means no output-token cost) are purpose-built for that use case. It’s not a general-purpose assistant and isn’t meant to replace conversational AI for open-ended tasks. It’s a narrower tool aimed squarely at automated decision pipelines.
Frequently Asked Questions #
What does “typed output” mean in Decision Node?
#
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
It means the model’s answer comes back as a specific data type your code can use directly, a boolean, a number, a probability, a label, or coordinates, rather than as a sentence of natural language that needs to be parsed or interpreted.
What are boosters in Decision Node?
Boosters are specialist models or data sources (such as age estimation, face detection, AI-image detection, OCR, or a custom knowledge base) that analyze the input first and pass their findings to the main model as evidence before it produces a final decision.
Does Decision Node generate any written text at all?
No. Every example shown returns structured values and confidence scores rather than written explanations, which is also why output tokens are described as free, since no text generation is happening on the output side.
How is confidence scoring useful in practice?
It tells a developer or an automated system how much to trust a given answer. A high-confidence result (like 0.99) can be auto-approved, while a low-confidence result (like 0.09 on a close call) can be routed to a human for review instead of acted on automatically.
Can Decision Node be run locally instead of through the API?
Open models related to Decision Node were mentioned as being released separately for local use, alongside the hosted API, CLI tool, and Python SDK options for accessing the main decision engine.