Hellixia
Hellixia is BayesiaLab’s Generative AI assistant. It turns a question, a document, or a set of nodes you already have into an explicit network: nodes, arcs, causal directions, comments, and probabilities that you can inspect, edit, and compute with.
The difference that matters is between generated text and a generated model. A Large Language Model asked about lung cancer risk returns prose. Hellixia asked the same question returns a Bayesian network whose nodes are propositions, whose arcs carry signed causal effects, and whose Conditional Probability Tables are already filled in, so inference can start the moment the network appears in the Graph Panel.
Every function lives in BayesiaLab’s Hellixia menu, and every result lands in the same modeling environment you use for learning, inference, effects analysis, simulation, and publishing. The Hellixia User Guide documents each function in operating detail.
What Hellixia Produces
Hellixia generates seven kinds of artifact. The first four are qualitative: they capture which concepts matter and how they relate, without effect sizes. The last three are quantified networks, ready for probabilistic inference.
| Artifact | What its nodes and arcs represent | Quantification |
|---|---|---|
| Knowledge Graph | Entities joined by labeled semantic relationships in seven families | None |
| Semantic Flowchart | Concepts, events, or steps joined by temporal dependency, meaning what must happen before what | None |
| Causal Semantic Diagram | Concepts joined by causal direction, mapping what causes what | None |
| Semantic Network | Concepts positioned by semantic proximity, with arcs learned from embeddings | Structure learned from embeddings |
| Constraint-Free Bayesian Network | Boolean, categorical, ordinal, or discretized continuous nodes | Full tabular Conditional Probability Tables |
| Propositional Causal Bayesian Network | Boolean propositions joined by signed direct effects from -1 to +1 | Tables written by the DualNoisyOr() or SwitchNoisyOr() function |
| Risk-Centric Causal Network | The same propositional form, with risk-management roles | Tables written by the same functions |
The qualitative graphs are not lesser versions of the quantified ones. They are the fastest way to map a domain before committing to probabilities, and each one can be reviewed, edited, and later promoted into a probabilistic model.
Generators, Finders, and Elicitors
Hellixia’s functions fall into three patterns, and recognizing which one you need is usually the quickest route to the right menu entry.
Generators build both nodes and arcs from scratch, either from the knowledge encoded in a Large Language Model or from a document you supply. You start with nothing but a topic.
Finders work on nodes that already exist in your graph. They add arcs and explanations among concepts you have defined, whether those nodes came from your own modeling, from imported data, or from an earlier Hellixia run.
Elicitors supply the numbers. Given a structure, they estimate prior probabilities for root nodes or local effects for each cause, so that a graph becomes a working probabilistic model.
From One Prompt to a Fully Specified Network
The Automatic Propositional Causal Bayesian Network Generator asks a single question: What would you like to model today?. The answer can be one concept, such as “Lung Cancer”, or a paragraph describing a domain.
Checking Advanced Mode opens the settings that shape the result. You choose the Completion Engine, optionally attach a Knowledge File, and add General Context to disambiguate the topic. Complexity guides how far upstream the network should reach, and Description Length controls how much text each node description carries. Risk-Centric Causal Network switches the node roles from Root Cause, Intermediate Cause, Intervention, Confounder, Main Criterion, and Effect to Threat, Preventive Barrier, Sub-Risk, Main Risk, Consequence, and Mitigating Barrier. Estimate Prior Probabilities for Root Nodes asks the model for the baseline rates of the risk factors it proposes.
The generated network places the Target Node at the center, with causes above and effects below. Every arc carries a signed direct effect, shown in blue when the cause promotes its effect and in red when it inhibits it, with the line thickness reflecting the magnitude. Arc comments state the mechanism and its weight, so a proposal such as “ongoing cigarette use increases inhaled carcinogens” arrives with a value of 92, while “effective cessation support reduces smoking-related exposure” arrives with -78.
Those two weights are the parameters of the formula that writes the child node’s probability table, which is why the model is complete on arrival: no cell is left blank, and no data is required. Root nodes carry the marginal probabilities the model estimated, for instance an 80 percent prevalence of current cigarette smoking, and node icons record where each number came from, whether retrieved from the language model, computed from a formula, or belonging to a latent node that summarizes several causes. Switching to Validation Mode starts inference immediately.
When Boolean propositions are too restrictive, the Automatic Bayesian Network Generator produces a constraint-free network instead. It decides for itself whether each concept is best represented as Boolean, categorical, ordinal, or continuous with automatic discretization, writes plain tabular probability tables rather than formula-driven ones, and groups the nodes into Classes.
Both generators produce a whole network in one pass. The Conversational Network Modeler takes the opposite approach, building the model through a dialogue in which you propose nodes, ask for the reasoning behind a suggested arc, and commit changes one at a time.
From Documents to Structure
Document Analysis routes one or more Knowledge Files to six generators. A Knowledge File can be a report, article, transcript, paper, or case study, in text, PDF, Word, Excel, RTF, HTML, XML, and PowerPoint formats, as well as audio recordings.
Five of the six generators are document-grounded counterparts of functions that otherwise draw on the model’s own knowledge: the Semantic Flowchart Generator, Causal Semantic Diagram Generator, Knowledge Graph Generator, Propositional Causal Bayesian Network Generator, and Semantic Network Generator. Each reads the file instead of, or in addition to, a typed prompt, which keeps the output anchored in a source you can check.
The sixth has no prompt-driven equivalent. The Doc-to-Node Generator creates one node per file and stores each document’s content as that node’s comment, so an entire corpus becomes a single network in which similar documents sit close to one another.
Semantic Structure from Embeddings
A Semantic Network is learned rather than written. It takes three steps: create the nodes, turn their text into embeddings, then learn which nodes belong near each other.
Nodes can be typed by hand, loaded from a data set, or proposed by the Dimension Elicitor, which probes a domain with keywords and creates a node for each dimension it finds. The Embedding Generator then reads each node’s name, long name, and comment and builds a vector of 1,024, 1,536, or 3,072 dimensions depending on the model, and it can concatenate vectors from several models to enrich the representation. Finally, unsupervised structural learning with the Maximum Weight Spanning Tree connects the nodes along their strongest semantic similarities.
Because the embedding is computed from a node’s text, the quality of that text decides the quality of the network. Richer comments produce more faithful proximity, which is why Hellixia’s comment tools are part of the semantic workflow and not merely cosmetic.
The Automatic Semantic Network Generator packages all three steps behind a single topic. You select the keywords that will probe the domain, either individually or through preselected groups such as Domain Description, Risk Analysis, Scientific article analysis, Book analysis, Ancestors, Descendants, and Pros & Cons, and it extracts the dimensions, computes the embeddings, synthesizes a data set, and learns the network.
Once a semantic network exists, Semantic Variable Clustering gathers nodes that mean similar things into classes, and the Class Description Generator replaces opaque cluster labels with descriptions of what each group actually represents.
Entities and Their Relationships
Semantic proximity says that two concepts are related. A Knowledge Graph says how. Hellixia classifies every relationship it proposes into one of seven families: taxonomic (a kidney stone is a calculus), partitive (a nephron is part of a kidney), attributive (urine has the property acidity), causal, functional (an NSAID is used for pain relief), temporal, and spatial (a calculus is located in the ureter).
The Knowledge Graph Generator builds such a graph from scratch, growing outward from a seed node. The Entity Relationship Finder does the same work in the other direction: give it a set of nodes, even nodes that carry nothing but a name, and it proposes the relationships among them along with the reasoning for each one.
Two generators specialize the relationship type. The Semantic Flowchart Generator keeps only temporal dependencies, which makes it the right tool for processes, pathways, and event chains where order carries the meaning. The Causal Semantic Diagram Generator keeps only causal links, producing an unquantified map of what causes what that serves as the structural precursor to a Causal Bayesian Network.
Causal Direction, Effects, and Priors
No method extracts causal direction from data alone. Given an association between two nodes, the data cannot say which one is the cause, so external knowledge has always been required, whether from a human expert or from the temporal order of the observations. Hellixia adds a third source by letting BayesiaLab query the domain knowledge held in a Large Language Model.
Pairwise Causal Link settles a single question, such as whether smoking causes lung cancer, and adds the arc together with a comment explaining the direction. The Causal Relationship Finder scales that to a whole selection, either examining every possible link or concentrating on the drivers of a chosen target, and it writes a node comment first for any node that lacks one so that the causal reasoning has something to work from. The Propositional Causal Bayesian Network Generator goes further and assembles the causal ancestors, and optionally the descendants, of a target node into a fully quantified model.
When a single opinion is not enough, the Multi-Engine Causal Relationship Finder puts every candidate relationship to several engines at once. Each returns a signed effect and a confidence level, the confidence levels are summed into an arc weight, and arcs are then added from the highest weight down, skipping any arc that would close a cycle. The result is robust to the quirks of any one model and is guaranteed to be a directed acyclic graph.
Numbers come from two elicitors. The Root Priors Elicitor estimates the marginal distribution of nodes that have no parents, which is their entire probability table. The ICI Local Effects Elicitor handles the harder case of a node with several parents by asking, for each parent in turn, what impact that cause has when it is the only active one and every other cause sits in its neutral state. A combination function then merges those local effects into the overall effect. The number of quantities to elicit therefore grows with the number of parents rather than with the product of their states, which is what makes large tables tractable for a model or an expert to fill.
LLM-Augmented Machine Learning
Hellixia does not have to replace machine learning. In two documented workflows it feeds BayesiaLab’s learning algorithms instead.
The first turns arcs into constraints. You learn a structure from data, select the arcs you want vetted, and Causal Structural Priors judges each relationship for causality, proposes a direction, and supplies a natural-language causal explanation you can accept or reject. Confirmed directions become structural priors that steer the next round of learning, which adds causal meaning to a structure that was originally only associative.
The second turns free text into variables. The Textual Dimension Elicitor reads the text columns of an imported data set, such as open-ended survey answers or product reviews, and creates a propositional node for each dimension those comments evoke. Textual Imputation then scores every comment against every proposition on a scale from 0 to 1 and writes those scores into the internal data set as new columns, discretized at 0.5 by default. From that point the text-derived nodes are ordinary variables, learnable by any BayesiaLab algorithm.
Documenting and Communicating a Model
A model that only its author can read is a model that cannot be reviewed. The Comments submenu holds seven tools that write node and arc comments from keywords or definitions, expand terse notes into readable explanations, condense long ones back to a tooltip, and derive a display label for a node from its comment.
Verbalize Relationships states in plain sentences what each arc means, so a clinician or an auditor can validate a relationship without reading a probability table. The Image Generator attaches an illustrative picture to each node for presentations, and the Translator renders node names, state names, and comments into another language through an engine such as ChatGPT or DeepL, leaving the structure untouched.
Setup, Context, and Review
Hellixia calls an external provider, so the first step is Settings, where you enter your provider account, API key, endpoint, and model choices. Completion and embedding models are configured separately, and a missing or expired key is the first thing to check when a function reports an error.
Two inputs do most of the work in steering a result. General Context resolves ambiguity that a short topic cannot, which matters whenever a term has several readings, as “Amazon” does for the river, the company, and the mythological warriors. A Knowledge File narrows the ground further by supplying the specific source material a general-purpose model would not have.
Everything Hellixia proposes is a hypothesis. Its value is that the hypothesis is explicit and auditable: arc comments record the claimed mechanism, causal explanations record the reasoning behind a direction, node icons record where each probability came from, and every element can be edited, learned over with data, or deleted. Confidential material sent to an external provider is subject to that provider’s terms, so check them and your own policy before running a function on sensitive content.
Publishing to HellixMap
Any network Hellixia generates can be published to HellixMap for browser-based exploration, sharing, and AI-assisted querying.
The two tools answer different questions. Hellixia is where a model is built, quantified, learned, and analyzed, inside BayesiaLab. HellixMap is where a finished model is navigated, published, and queried, in a browser and without a BayesiaLab license.
Examples
Hellixia’s output is easiest to judge on real material. The examples cover philosophy, literature, song lyrics, cinema, animals, and causal modeling, alongside step-by-step tutorials and webinar workflows.