AI Functions

AI-powered text processing using embeddings and large language models.

Vector Embeddings

Generate and compare vector embeddings for semantic similarity:

embeddings:vectorize(arg1: string, arg2: string, …) → array<double>
embeddings:cosine_sim_udf(left: array<double>, right: array<double>) → double
embeddings:cosine_sim(left: array<double>, right: array<double>) → double

Example:

% Generate embeddings for similarity analysis
customer_similarity(Name1, Name2, Similarity) <- 
    customer(Name1, Info1), 
    customer(Name2, Info2), 
    Vec1 = embeddings:vectorize(Name1, Info1), 
    Vec2 = embeddings:vectorize(Name2, Info2), 
    Similarity = embeddings:cosine_sim(Vec1, Vec2).

LLM Generation

Generate dynamic content using large language models. Prometheux provides two approaches:

llm:generate - approach for simple use cases
#LLM - approach for large-scale processing

Approach 1: `llm:generate`

Traditional UDF for straightforward use cases with automatic LLM pool support:

llm:generate(prompt: string) → string
llm:generate(prompt: string, options: string) → typed
llm:generate(prompt: string, options: string, arg1: any, arg2: any, ...) → typed

Options Format (second argument, optional, comma-separated):

"output_type=string" - Specify output type (default: string)
"selected_models=gpt-4o" - Filter to single model type
"selected_models=gpt-4o;gpt-4o-mini;gpt-4.1" - Filter to multiple model types (semicolon-separated)
"output_type=int,selected_models=gpt-4o;gpt-4o-mini" - Combine output type with multiple models

Flexible Second Argument: Can be omitted or used as:

Omitted → llm:generate(prompt, dataArg1, ...) - Defaults to string output, all models
Options string → llm:generate(prompt, "output_type=int,selected_models=...") - Parse as configuration
Data argument → llm:generate(prompt, dataArg1, dataArg2) - If not an options string, treated as first data arg

Supported Output Types: string, int, double, boolean, list<string>, list<int>, list<double>, list<boolean>

Prompt Templating: Use ${Variable} for direct variable interpolation or ${arg_1}, ${arg_2}, etc. for positional arguments.

Model Selection: Use semicolon-separated list in selected_models to filter which models to use:

Single model: "selected_models=gpt-4o"
Multiple models: "selected_models=gpt-4o;gpt-4o-mini;gpt-4.1;gpt-4.1-nano"
If not specified, all configured pools are used automatically for load balancing.

Examples:

Simplest usage (defaults to string output):

% Ask a question with default string output
answer(Q, A) <- 
    question(Q), 
    A = llm:generate(Q).

% Generate description with variable interpolation
product_summary(Name, Price, Summary) <- 
    product(Name, Price), 
    Summary = llm:generate("Describe ${Name} priced at ${Price}").

Using options:

% Specify output type only
count(Data, Num) <- 
    data(Data), 
    Num = llm:generate("Count items in: ${Data}", "output_type=int").

% Specify model selection only (defaults to string)
fast_answer(Q, A) <- 
    question(Q), 
    A = llm:generate(Q, "selected_models=gpt-4o-mini").

% Combine both options
analysis(Data, Result) <- 
    data(Data), 
    Result = llm:generate("Analyze: ${Data}", "output_type=boolean,selected_models=gpt-4o").

% Use multiple models for load balancing
balanced_query(Q, A) <- 
    question(Q), 
    A = llm:generate(Q, "selected_models=gpt-4o;gpt-4o-mini;gpt-4.1").

% Skip options entirely and pass data args directly (defaults to string, all models)
description(Name, Age, Desc) <- 
    person(Name, Age), 
    Desc = llm:generate("Describe {arg_1} aged {arg_2}", Name, Age).

Using positional arguments with options format:

% Generate with arg_1, arg_2 placeholders and explicit output types
diagnose_with_args(PatientID, HasDiagnosis, Explanation) <- 
    clinical_notes(PatientID, Notes), 
    HasDiagnosis = llm:generate(
        "Review the clinical notes for patient ${arg_1}: ${arg_2}. Determine if there is a diagnosis.",
        "output_type=boolean",
        PatientID,
        Notes
    ),
    Explanation = llm:generate(
        "Analyze the notes for patient ${arg_1}: ${arg_2}. Provide a brief rationale.",
        "output_type=string",
        PatientID,
        Notes
    ).

Multiple output types:

% String output for descriptions (default)
product_description(Product, Description) <- 
    product(Product, Features, Price), 
    Description = llm:generate(
        "Create a marketing description for ${Product} with features ${Features} priced at ${Price}"
    ).

% Boolean output for classification
is_positive_feedback(FeedbackID, IsPositive) <- 
    feedback(FeedbackID, Text), 
    IsPositive = llm:generate(
        "Is this feedback positive? ${Text}",
        "output_type=boolean"
    ).

Model selection for load balancing and quality control:

% Use only gpt-4o models for critical analysis (comma-separated options)
critical_analysis(PatientID, Diagnosis) <- 
    patient(PatientID, Symptoms), 
    Diagnosis = llm:generate(
        "Based on symptoms ${arg_1}, provide a preliminary diagnosis.",
        "output_type=string,selected_models=gpt-4o",
        Symptoms
    ).

% Use fast models only (defaults to string output)
quick_category(Doc, Category) <- 
    document(Doc, Text), 
    Category = llm:generate(
        "Classify this document into one category: ${Text}",
        "selected_models=gpt-4o-mini"
    ).

% Specify both output type and models explicitly
urgent_check(Text, IsUrgent) <- 
    feedback(Text), 
    IsUrgent = llm:generate(
        "Is this urgent? ${Text}",
        "output_type=boolean,selected_models=gpt-4o"
    ).

% Balance across multiple model types with data arguments
balanced_processing(Item, Result) <- 
    items(Item, Data), 
    Result = llm:generate(
        "Process this item: ${arg_1}",
        "selected_models=gpt-4o;gpt-4o-mini;gpt-4.1",
        Data
    ).

% Large-scale processing with many model types
large_scale(Data, Analysis) <- 
    dataset(Data), 
    Analysis = llm:generate(
        "Analyze: ${Data}",
        "selected_models=gpt-4o;gpt-4o-mini;gpt-4.1;gpt-4.1-nano;gpt-4o-nano"
    ).

When to use selected_models with llm:generate:

Quality control: Route critical queries to premium models (gpt-5-nano)
Performance optimization: Use faster models (gpt-4o-mini,gpt-4o,gpt-4.1,gpt-4.1-nano,gpt-4.1-mini) for simple tasks
Cost optimization: Balance cost and quality across model tiers
Load balancing: Distribute work across specific model types
Model testing: Test specific models without changing system configuration

If selected_models is not specified, all configured LLM pools are used automatically for optimal load distribution.

Approach 2: `#LLM`

For large-scale processing across LLM endpoints:

#LLM(input_relation, parameters)

The #LLM function processes all rows from the input relation in parallel, distributing work across multiple LLM endpoints for optimal performance.

Parameters

All parameters are specified as a comma-separated string:

Required:

prompt=<text> - Single prompt template
OR prompt_1=<text1>,prompt_2=<text2>,... - Multiple numbered prompts
OR prompts=prompt_1:<text1>;prompt_2:<text2>;... - Compact format for multiple prompts

Optional:

output_type=<type> - Output type for all prompts (default: string)
output_type_1=<type1>,output_type_2=<type2>,... - Per-prompt output types (overrides output_type)
projected_columns=<spec> - Custom output column ordering (e.g., arg_1:llm_2:llm_1:arg_3)
selected_models=<models> - Semicolon-separated list of models to use (e.g., gpt-4o;gpt-4o-mini)
num_partitions=<N> - Manual partition override

Supported Output Types: string, int, integer, long, double, float, number, boolean, bool, list<string>, list<int>, list<double>, set<string>

Prompt Templating: Use {arg_1}, {arg_2}, etc. to reference input columns (1-based indexing).

Model Selection with `selected_models`

By default, #LLM uses all configured LLM endpoints. Use selected_models to filter which model types to use:

% Use only gpt-4o endpoints (skip slower/experimental models)
answer(Q, A) <- 
    #LLM(question, "prompt=Answer: {arg_1},selected_models=gpt-4o").

% Use multiple model types
fast_analysis(Q, A) <- 
    #LLM(question, "prompt=Analyze: {arg_1},selected_models=gpt-4o;gpt-4o-mini").

Supported models: gpt-4o, gpt-4o-mini, gpt-5-nano, gpt-5-mini, gpt-4.1-nano, gpt-4.1-mini, and others configured in your environment.

Use cases:

Performance optimization: Route simple queries to faster models (gpt-4o-mini)
Quality control: Use only high-quality models (gpt-4o) for critical analysis
Cost optimization: Balance speed and cost across model tiers
Model testing: Test specific models without changing configuration

If selected_models is not specified, all configured endpoints are used for maximum parallelization.

Why `num_partitions` is Optional

Prometheux automatically handles optimal partitioning based on your data source, especially if it is already partitioned. Only specify num_partitions if you need to override the automatic behavior for specific performance requirements.

`#LLM` Function Examples

Example 1: Single Prompt with Default Output

Process customer feedback with automatic output type (defaults to string):

% Classify customer sentiment
sentiment_analysis(FeedbackID, Text, Sentiment) <- 
    #LLM(feedback, "prompt=Classify the sentiment of this feedback as positive, neutral, or negative: {arg_2}").

% Usage with input data:
feedback(1, "Great product, very satisfied!").
feedback(2, "Terrible experience, would not recommend.").

Output:

Input columns: feedback_0 (FeedbackID), feedback_1 (Text)
Output columns: feedback_0, feedback_1, feedback_2 (Sentiment result)

Example 2: Multiple Prompts with Different Output Types

Generate multiple insights with type-specific outputs:

% Analyze product reviews with multiple outputs
review_analysis(ProductID, Review, Sentiment, Rating, Keywords) <- 
    #LLM(reviews, 
        "prompt_1=What is the sentiment: {arg_2}?,
         prompt_2=Rate this review from 1-10 (just the number): {arg_2},
         prompt_3=Extract key topics from: {arg_2},
         output_type_1=string,
         output_type_2=int,
         output_type_3=string").

% Input data:
reviews(101, "Amazing quality but expensive").

Output:

reviews_0: ProductID (101)
reviews_1: Review text
reviews_2: Sentiment (string) from prompt_1
reviews_3: Rating (integer) from prompt_2
reviews_4: Keywords (string) from prompt_3

Example 3: Custom Column Ordering with `projected_columns`

Reorder output columns to match your head predicate:

% Place LLM results before input columns
enriched_data(Summary, OriginalText, WordCount) <- 
    #LLM(documents, 
        "prompt_1=Summarize in one sentence: {arg_1},
         prompt_2=Count words in: {arg_1} (answer with just the number),
         output_type_1=string,
         output_type_2=int,
         projected_columns=llm_1:arg_1:llm_2").

% Input:
documents("Artificial intelligence is transforming healthcare.").

Output column order:

enriched_0: Summary (llm_1)
enriched_1: Original text (arg_1)
enriched_2: Word count (llm_2)

Example 4: Compact Format with Single Output Type

Use compact prompts format when all outputs share the same type:

% Multiple text analyses with single output_type
text_processing(Text, Translation, Paraphrase, Summary) <- 
    #LLM(sentences, 
        "prompts=prompt_1:Translate to French: {arg_1};
                 prompt_2:Paraphrase this: {arg_1};
                 prompt_3:Summarize this: {arg_1},
         output_type=string").

% Input:
sentences("Machine learning enables computers to learn from data.").

Output:

All three LLM results are strings
output_type=string applies to all prompts (prompt_1, prompt_2, prompt_3)

Example 5: Model Selection for Optimal Performance

Choose specific models based on your quality, speed, and cost requirements:

% High-quality analysis: Use only gpt-4o
detailed_analysis(Doc, Analysis) <- 
    #LLM(documents, 
        "prompt=Provide a detailed analysis of: {arg_1},
         output_type=string,
         selected_models=gpt-4o").

% Fast classification: Use faster models for simple tasks
quick_category(Text, Category) <- 
    #LLM(items, 
        "prompt=Classify this into one category: {arg_1},
         output_type=string,
         selected_models=gpt-4o-mini").

% Balanced approach: Use multiple model types
balanced_processing(Data, Result) <- 
    #LLM(dataset, 
        "prompt=Process: {arg_1},
         output_type=string,
         selected_models=gpt-4o;gpt-4o-mini").

Why use selected_models:

Filter out slower or experimental models (e.g., exclude gpt-5-nano preview models)
Route high-priority queries to premium models (gpt-4o)
Route high-volume queries to cost-effective models (gpt-4o-mini)
Test specific models without changing system configuration

Example 6: Processing Large Datasets (Advanced)

For specific performance tuning on very large datasets, you can override automatic partitioning:

% Process millions of records with explicit partition control
batch_classification(DocID, Text, Category) <- 
    #LLM(documents, 
        "prompt=Categorize this document: {arg_2},
         output_type=string,
         num_partitions=32").

% Note: Prometheux automatically optimizes partitioning
% Only specify num_partitions for advanced performance tuning

When to use num_partitions:

Very large datasets where you need explicit control over parallelism
Performance tuning and optimization scenarios
Default automatic partitioning is optimal for most use cases

Use Cases

Semantic Search

Combine embeddings with LLM for intelligent document retrieval:

% Find and summarize similar documents
similar_documents(Doc1, Doc2, Score, Summary) <- 
    document(Doc1, Content1), 
    document(Doc2, Content2), 
    Vec1 = embeddings:vectorize(Content1), 
    Vec2 = embeddings:vectorize(Content2), 
    Score = embeddings:cosine_sim(Vec1, Vec2), 
    Score > 0.8,
    result(_, _, Summary) <- #LLM(document, "prompt=Summarize the key similarities between these documents: {arg_2},output_type=string").

Content Classification

Classify large volumes of text in parallel with optimized model selection:

% Classify customer feedback at scale using fast models
classify_feedback(FeedbackID, Text, Category, IsUrgent) <- 
    #LLM(feedback, 
        "prompt_1=Classify this feedback as positive, neutral, or negative: {arg_2},
         prompt_2=Is this feedback urgent? Answer yes or no: {arg_2},
         output_type_1=string,
         output_type_2=string,
         selected_models=gpt-4o;gpt-4o-mini").

% Input data:
feedback(1, "Excellent service, very happy!").
feedback(2, "URGENT: Product stopped working immediately").

Data Enrichment

Generate missing information for entire datasets with quality-focused model selection:

% Enrich product catalog with AI-generated content using high-quality models
enrich_product(ProductID, Name, Category, Description, TargetAge, Keywords) <- 
    #LLM(product, 
        "prompt_1=Generate a marketing description for {arg_2} in {arg_3} category,
         prompt_2=What age group is {arg_2} best suited for? Answer with a number (e.g., 25),
         prompt_3=List 3 keywords for {arg_2},
         output_type_1=string,
         output_type_2=int,
         output_type_3=string,
         selected_models=gpt-4o").

% Automatically processes entire product catalog in parallel
product(101, "Smart Watch Pro", "Electronics").
product(102, "Yoga Mat Premium", "Fitness").
product(103, "Coffee Maker Deluxe", "Home Appliances").

Healthcare Decision Support

Multi-faceted analysis of clinical data using high-quality models for critical healthcare decisions:

% Analyze patient records with multiple AI insights using only gpt-4o
patient_analysis(PatientID, Notes, RiskLevel, Recommendations, FollowUpDays) <- 
    #LLM(clinical_notes, 
        "prompt_1=Assess risk level (low/medium/high) for: {arg_2},
         prompt_2=Provide 3 key recommendations for: {arg_2},
         prompt_3=How many days until follow-up? Answer with just a number for: {arg_2},
         output_type_1=string,
         output_type_2=string,
         output_type_3=int,
         projected_columns=arg_1:arg_2:llm_1:llm_2:llm_3,
         selected_models=gpt-4o").

clinical_notes(12345, "Patient presents with elevated blood pressure...").

Note: For healthcare and other critical applications, use selected_models=gpt-4o to ensure the highest quality and reliability.

Vector Embeddings​

LLM Generation​

Approach 1: llm:generate​

Approach 2: #LLM​

Parameters​

Model Selection with selected_models​

Why num_partitions is Optional​

#LLM Function Examples​

Example 1: Single Prompt with Default Output​

Example 2: Multiple Prompts with Different Output Types​

Example 3: Custom Column Ordering with projected_columns​

Example 4: Compact Format with Single Output Type​

Example 5: Model Selection for Optimal Performance​

Example 6: Processing Large Datasets (Advanced)​

Use Cases​

Semantic Search​

Content Classification​

Data Enrichment​

Healthcare Decision Support​

Vector Embeddings

LLM Generation

Approach 1: `llm:generate`

Approach 2: `#LLM`

Parameters

Model Selection with `selected_models`

Why `num_partitions` is Optional

`#LLM` Function Examples

Example 1: Single Prompt with Default Output

Example 2: Multiple Prompts with Different Output Types

Example 3: Custom Column Ordering with `projected_columns`

Example 4: Compact Format with Single Output Type

Example 5: Model Selection for Optimal Performance

Example 6: Processing Large Datasets (Advanced)

Use Cases

Semantic Search

Content Classification

Data Enrichment

Healthcare Decision Support