Orb

Documentation

Tensors, neural nets and LLMs

Every operator a .lolo program can call, from the canonical registry the compiler bakes.

On this page

Namespaces on this page: llm/, nn/, tensor/, train/.

llm/

`llm/call-tools`

Call the LLM with tool definitions. Returns the assistant response with optional tool calls and token usage.

lolo

(llm/call-tools prompt tools)

Parameters

Name	Type	Description
`prompt`	`string`	User prompt
`tools`	`array`	Tool definitions

Returns: LlmCallToolsResult · Runs on: any

Example

lolo

(llm/call-tools [{"role":"user","content":"What time is it?"}] [{"name":"getTime","description":"Get current time","parameters":{}}])

`llm/compact`

Compact the context window. Returns before/after token counts.

lolo

(llm/compact strategy)

Parameters

Name	Type	Description
`strategy`	`{'kind': 'union', 'of': [{'kind': 'literal', 'value': 'hybrid'}, {'kind': 'literal', 'value': 'summarize'}, {'kind': 'literal', 'value': 'truncate'}, {'kind': 'literal', 'value': 'extract'}]}`	Compaction strategy

Returns: object · Runs on: any

Example

lolo

(llm/compact) // => {before: 10000 after: 8000}

`llm/embed`

Generate embeddings for an array of texts. Returns a 2D array of floats.

lolo

(llm/embed texts)

Parameters

Name	Type	Description
`texts`	`array`	Texts to embed

Returns: array · Runs on: any

Example

lolo

(llm/embed ["hello" "world"])

`llm/generate`

Generate text from an LLM. Returns the generated string.

lolo

(llm/generate prompt)

Parameters

Name	Type	Description
`prompt`	`string`	Generation prompt

Returns: string · Runs on: any

Example

lolo

(llm/generate "Write a haiku") // => "Spring rain falls gently"

`llm/switch`

Switch the active LLM provider and optionally the model.

lolo

(llm/switch provider)

Parameters

Name	Type	Description
`provider`	`string`	LLM provider name

Returns: void · Runs on: any

Example

lolo

(llm/switch "claude")

`llm/token-count`

Get the current token count in the context window.

lolo

(llm/token-count)

Returns: number

Example

lolo

(llm/token-count) // => 5432

nn/

`nn/batchnorm`

Batch normalization layer

lolo

(nn/batchnorm input)

Parameters

Name	Type	Description
`input`	`any`	Input layer or config

Returns: nn/layer

Example

lolo

(nn/batchnorm 64)

`nn/clone`

Create a deep copy of the network with same weights

lolo

(nn/clone module)

Parameters

Name	Type	Description
`module`	`any`	Neural network module

Returns: nn/module

Example

lolo

(nn/clone {"layers":[{"type":"dense","units":2}]})

`nn/dropout`

Dropout layer for regularization (randomly zeros elements during training)

lolo

(nn/dropout p)

Parameters

Name	Type	Description
`p`	`number`	Dropout probability

Returns: nn/layer

Example

lolo

(nn/dropout)

`nn/forward`

Execute forward pass through network

lolo

(nn/forward module input)

Parameters

Name	Type	Description
`module`	`any`	Neural network module
`input`	`any`	Input tensor

Returns: tensor

Example

lolo

(nn/forward {"layers":[{"type":"dense","units":2}]} [0.1,0.2])

`nn/getWeights`

Get network weights as a flat tensor

lolo

(nn/getWeights module)

Parameters

Name	Type	Description
`module`	`any`	Neural network module

Returns: tensor

Example

lolo

(nn/getWeights {"layers":[{"type":"dense","units":2}]})

`nn/layernorm`

Layer normalization

lolo

(nn/layernorm input)

Parameters

Name	Type	Description
`input`	`any`	Input layer or config

Returns: nn/layer

Example

lolo

(nn/layernorm 64)

`nn/linear`

Fully connected linear layer (output = input * weights + bias)

lolo

(nn/linear inSize outSize)

Parameters

Name	Type	Description
`inSize`	`number`	Input dimension
`outSize`	`number`	Output dimension

Returns: nn/layer

Example

lolo

(nn/linear 128 64)

`nn/paramCount`

Get total number of trainable parameters

lolo

(nn/paramCount module)

Parameters

Name	Type	Description
`module`	`any`	Neural network module

Returns: number

Example

lolo

(nn/paramCount {"layers":[{"type":"dense","units":2}]})

`nn/relu`

ReLU activation function: max(0, x)

lolo

(nn/relu)

Returns: nn/layer

Example

lolo

(nn/relu)

`nn/sequential`

Create a sequential neural network from layers

lolo

(nn/sequential layer)

Parameters

Name	Type	Description
`layer`	`any`	First layer (variadic)

Returns: nn/module

Example

lolo

(nn/sequential ["nn/linear",16,64] ["nn/relu"] ["nn/linear",64,4])

`nn/setWeights`

Set network weights from a flat tensor

lolo

(nn/setWeights module weights)

Parameters

Name	Type	Description
`module`	`any`	Neural network module
`weights`	`any`	Weight tensor (variadic)

Returns: nn/module · Runs on: any

Example

lolo

(nn/setWeights {"layers":[{"type":"dense","units":2}]} [0.1,0.2,0.3])

`nn/sigmoid`

Sigmoid activation function: 1 / (1 + e^-x)

lolo

(nn/sigmoid)

Returns: nn/layer

Example

lolo

(nn/sigmoid)

`nn/softmax`

Softmax activation function (normalizes to probability distribution)

lolo

(nn/softmax dim)

Parameters

Name	Type	Description
`dim`	`number`	Dimension to apply softmax

Returns: nn/layer

Example

lolo

(nn/softmax)

`nn/tanh`

Tanh activation function: (e^x - e^-x) / (e^x + e^-x)

lolo

(nn/tanh)

Returns: nn/layer

Example

lolo

(nn/tanh)

tensor/

`tensor/add`

Element-wise addition

lolo

(tensor/add a b)

Parameters

Name	Type	Description
`a`	`any`	First tensor
`b`	`any`	Second tensor

Returns: tensor

Example

lolo

(tensor/add [1,2] [3,4])

`tensor/allInRange`

Check if all elements are within range [min, max]

lolo

(tensor/allInRange tensor range)

Parameters

Name	Type	Description
`tensor`	`any`	Tensor to check
`range`	`array`	Range [min, max]

Returns: boolean

Example

lolo

(tensor/allInRange [0.5,-0.2] [-1.0,1.0]) // => true

`tensor/argmax`

Index of maximum value

lolo

(tensor/argmax tensor)

Parameters

Name	Type	Description
`tensor`	`any`	Tensor to search (variadic)

Returns: number | tensor

Example

lolo

(tensor/argmax [0.1,0.9,0.3]) // => 1

`tensor/cat`

Concatenate two tensors along the last dimension

lolo

(tensor/cat a b)

Parameters

Name	Type	Description
`a`	`tensor`	First tensor
`b`	`tensor`	Second tensor

Returns: tensor

Example

lolo

(tensor/cat [1,2] [3,4])

`tensor/clamp`

Clamp all elements to range [min, max]

lolo

(tensor/clamp t min max)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor to clamp
`min`	`number`	Minimum value
`max`	`number`	Maximum value

Returns: tensor

Example

lolo

(tensor/clamp [5,-15,3] -10.0 10.0)

`tensor/clampPerDim`

Clamp each dimension to its specified range

lolo

(tensor/clampPerDim t ranges)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor to clamp
`ranges`	`object`	Per-dimension range specifications

Returns: tensor

Example

lolo

(tensor/clampPerDim [5,-15] {"0":{"min":0,"max":10},"1":{"min":-10,"max":10}})

`tensor/div`

Element-wise division

lolo

(tensor/div a b)

Parameters

Name	Type	Description
`a`	`tensor`	Dividend tensor
`b`	`tensor`	Divisor tensor

Returns: tensor

Example

lolo

(tensor/div [10,20] 2)

`tensor/dot`

Dot product of two 1D tensors

lolo

(tensor/dot a b)

Parameters

Name	Type	Description
`a`	`tensor`	First 1D tensor
`b`	`tensor`	Second 1D tensor

Returns: number

Example

lolo

(tensor/dot [1,2] [3,4]) // => 11

`tensor/flatten`

Flatten tensor to 1D

lolo

(tensor/flatten t)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor to flatten

Returns: tensor

Example

lolo

(tensor/flatten [[1,2],[3,4]])

`tensor/from`

Create tensor from array

lolo

(tensor/from arr)

Parameters

Name	Type	Description
`arr`	`array`	Array to convert

Returns: tensor

Example

lolo

(tensor/from [1 2 3 4])

`tensor/get`

Get element at index

lolo

(tensor/get t idx)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor to index
`idx`	`number`	Index position

Returns: number

Example

lolo

(tensor/get [10,20,30,40] 3) // => 3.14

`tensor/matmul`

Matrix multiplication

lolo

(tensor/matmul a b)

Parameters

Name	Type	Description
`a`	`tensor`	First matrix
`b`	`tensor`	Second matrix

Returns: tensor

Example

lolo

(tensor/matmul [[1,2],[3,4]] [[5,6],[7,8]])

`tensor/max`

Maximum value in tensor

lolo

(tensor/max t)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor

Returns: number | tensor

Example

lolo

(tensor/max [0.1,0.9,0.3]) // => 99

`tensor/mean`

Mean of tensor elements

lolo

(tensor/mean t)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor

Returns: number | tensor

Example

lolo

(tensor/mean [0.2,0.4,0.6]) // => 50

`tensor/min`

Minimum value in tensor

lolo

(tensor/min t)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor

Returns: number | tensor

Example

lolo

(tensor/min [3,1,2]) // => 1

`tensor/mul`

Element-wise multiplication

lolo

(tensor/mul a b)

Parameters

Name	Type	Description
`a`	`tensor`	First tensor
`b`	`tensor`	Second tensor

Returns: tensor

Example

lolo

(tensor/mul [0.5,0.7] 0.99)

`tensor/norm`

L2 norm of tensor

lolo

(tensor/norm t)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor

Returns: number

Example

lolo

(tensor/norm [3,4]) // => 5.2

`tensor/ones`

Create tensor filled with ones

lolo

(tensor/ones shape)

Parameters

Name	Type	Description
`shape`	`array`	Tensor shape

Returns: tensor

Example

lolo

(tensor/ones [2 3])

`tensor/outOfRangeDims`

Get dimensions that exceed their specified ranges

lolo

(tensor/outOfRangeDims t ranges)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor to check
`ranges`	`object`	Dimension range specifications

Returns: array

Example

lolo

(tensor/outOfRangeDims [5,-15] {"0":{"min":0,"max":10},"1":{"min":-10,"max":10}}) // => [0]

`tensor/outOfRangeIndices`

Get indices of elements outside range

lolo

(tensor/outOfRangeIndices t ranges)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor to check
`ranges`	`object`	Value range specifications

Returns: number[]

Example

lolo

(tensor/outOfRangeIndices [0.5,1.5] [-1.0,1.0]) // => [0 2]

`tensor/rand`

Create tensor with random values in [0, 1)

lolo

(tensor/rand shape)

Parameters

Name	Type	Description
`shape`	`array`	Tensor shape

Returns: tensor

Example

lolo

(tensor/rand [2 3])

`tensor/randn`

Create tensor with random values from standard normal distribution

lolo

(tensor/randn shape)

Parameters

Name	Type	Description
`shape`	`array`	Tensor shape

Returns: tensor

Example

lolo

(tensor/randn [2 3])

`tensor/reshape`

Reshape tensor to new shape (total elements must match)

lolo

(tensor/reshape t shape)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor to reshape
`shape`	`array`	New shape

Returns: tensor

Example

lolo

(tensor/reshape [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16] [4,4])

`tensor/shape`

Get tensor shape as array

lolo

(tensor/shape t)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor

Returns: number[]

Example

lolo

(tensor/shape [[1,2],[3,4]]) // => [3 4 5]

`tensor/slice`

Get slice of tensor

lolo

(tensor/slice t start end)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor to slice
`start`	`number`	Start index
`end`	`number`	End index

Returns: tensor

Example

lolo

(tensor/slice [10,20,30,40] 0 3)

`tensor/squeeze`

Remove all dimensions of size one

lolo

(tensor/squeeze t)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor to squeeze

Returns: tensor

Example

lolo

(tensor/squeeze [[1,2,3]])

`tensor/stack`

Stack two tensors along a new leading dimension

lolo

(tensor/stack a b)

Parameters

Name	Type	Description
`a`	`tensor`	First tensor
`b`	`tensor`	Second tensor

Returns: tensor

Example

lolo

(tensor/stack [1,2] [3,4])

`tensor/sub`

Element-wise subtraction

lolo

(tensor/sub a b)

Parameters

Name	Type	Description
`a`	`tensor`	Minuend tensor
`b`	`tensor`	Subtrahend tensor

Returns: tensor

Example

lolo

(tensor/sub [1,2] [3,4])

`tensor/sum`

Sum of tensor elements

lolo

(tensor/sum t)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor

Returns: number | tensor

Example

lolo

(tensor/sum [1,0,1]) // => 500

`tensor/toArray`

Convert tensor to nested array

lolo

(tensor/toArray t)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor to convert

Returns: array

Example

lolo

(tensor/toArray [[1,2],[3,4]]) // => [[1 2] [3 4]]

`tensor/toList`

Convert 1D tensor to flat array

lolo

(tensor/toList t)

Parameters

Name	Type	Description
`t`	`tensor`	1D tensor

Returns: number[]

Example

lolo

(tensor/toList [1,2,3]) // => [1 2 3]

`tensor/transpose`

Transpose a tensor (optionally swapping two named dimensions)

lolo

(tensor/transpose t)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor to transpose

Returns: tensor

Example

lolo

(tensor/transpose [[1,2],[3,4]] 0 1)

`tensor/unsqueeze`

Insert a dimension of size one at the given position

lolo

(tensor/unsqueeze t dim)

Parameters

Name	Type	Description
`t`	`tensor`	Tensor to unsqueeze
`dim`	`number`	Dimension position

Returns: tensor

Example

lolo

(tensor/unsqueeze [1,2,3] 0)

`tensor/zeros`

Create tensor filled with zeros

lolo

(tensor/zeros shape)

Parameters

Name	Type	Description
`shape`	`array`	Tensor shape

Returns: tensor

Example

lolo

(tensor/zeros [2 3])

train/

`train/adam`

Adam optimizer step

lolo

(train/adam params grads)

Parameters

Name	Type	Description
`params`	`object`	Model parameters
`grads`	`object`	Gradient values

Returns: void · Runs on: any

Example

lolo

(train/adam {"layers":[{"type":"dense","units":2}]} 0.001)

`train/checkConstraints`

Check if network weights satisfy all constraints

lolo

(train/checkConstraints weights constraints)

Parameters

Name	Type	Description
`weights`	`object`	Network weights
`constraints`	`object`	Constraint specifications

Returns: train/constraintResult

Example

lolo

(train/checkConstraints [0.1,0.2] {"maxMagnitude":10})

`train/checkForbiddenOutputs`

Check if model produces outputs in forbidden regions

lolo

(train/checkForbiddenOutputs model outputs ranges)

Parameters

Name	Type	Description
`model`	`object`	Model
`outputs`	`array`	Model outputs
`ranges`	`object`	Forbidden output ranges

Returns: train/forbiddenResult

Example

lolo

(train/checkForbiddenOutputs {"layers":[{"type":"dense","units":2}]} [[0],[1]] {"0":{"min":100,"max":200}})

`train/checkRegression`

Check if new model regresses on required invariants

lolo

(train/checkRegression oldModel newModel testData)

Parameters

Name	Type	Description
`oldModel`	`object`	Old model
`newModel`	`object`	New model
`testData`	`array`	Test dataset

Returns: train/regressionResult

Example

lolo

(train/checkRegression [0.1,0.2] {"layers":[{"type":"dense","units":2}]} [{"input":[0],"expected":[1]}])

`train/checkWeightMagnitude`

Check if all weights are within magnitude limit

lolo

(train/checkWeightMagnitude weights maxMag)

Parameters

Name	Type	Description
`weights`	`object`	Network weights
`maxMag`	`number`	Maximum magnitude

Returns: boolean

Example

lolo

(train/checkWeightMagnitude {"layers":[{"type":"dense","units":2}]} 10.0) // => true

`train/clipGradients`

Clip gradients to max norm (modifies in place)

lolo

(train/clipGradients grads maxNorm)

Parameters

Name	Type	Description
`grads`	`object`	Gradients to clip
`maxNorm`	`number`	Maximum norm

Returns: number · Runs on: any

Example

lolo

(train/clipGradients {"layers":[{"type":"dense","units":2}]} 1.0)

`train/clipWeights`

Clip weights to max magnitude (modifies in place)

lolo

(train/clipWeights weights maxMag)

Parameters

Name	Type	Description
`weights`	`object`	Weights to clip
`maxMag`	`number`	Maximum magnitude

Returns: void · Runs on: any

Example

lolo

(train/clipWeights {"layers":[{"type":"dense","units":2}]} 10.0)

`train/computeAdvantages`

Compute GAE advantages for policy gradient

lolo

(train/computeAdvantages rewards baseline gamma)

Parameters

Name	Type	Description
`rewards`	`array`	Reward trajectory
`baseline`	`array`	Baseline values
`gamma`	`number`	Discount factor

Returns: tensor

Example

lolo

(train/computeAdvantages [1,0,1] [0.5,0.4,0.3] {"gamma":0.99,"lambda":0.95})

`train/computeReturns`

Compute discounted returns from rewards

lolo

(train/computeReturns rewards gamma)

Parameters

Name	Type	Description
`rewards`	`array`	Reward trajectory
`gamma`	`number`	Discount factor

Returns: tensor

Example

lolo

(train/computeReturns [1,0,1] 0.99)

`train/crossEntropy`

Cross-entropy loss for classification

lolo

(train/crossEntropy predictions targets)

Parameters

Name	Type	Description
`predictions`	`array`	Predicted probabilities
`targets`	`array`	Target labels

Returns: number

Example

lolo

(train/crossEntropy [0.1 0.9] [0 1]) // => 0.105

`train/getGradientNorm`

Get current gradient norm

lolo

(train/getGradientNorm gradients)

Parameters

Name	Type	Description
`gradients`	`object`	Gradient structure

Returns: number

Example

lolo

(train/getGradientNorm {"layers":[{"type":"dense","units":2}]})

`train/getMaxWeightMagnitude`

Get maximum weight magnitude in network

lolo

(train/getMaxWeightMagnitude model)

Parameters

Name	Type	Description
`model`	`object`	Neural network model

Returns: number

Example

lolo

(train/getMaxWeightMagnitude {"layers":[{"type":"dense","units":2}]})

`train/huber`

Huber loss (smooth L1, robust to outliers)

lolo

(train/huber predicted target)

Parameters

Name	Type	Description
`predicted`	`number`	Predicted value
`target`	`number`	Target value (variadic)

Returns: number

Example

lolo

(train/huber 0.5 0.7 1.0)

`train/loop`

Execute training loop with constraints

lolo

(train/loop model config callback)

Parameters

Name	Type	Description
`model`	`object`	Neural network model
`config`	`object`	Training configuration
`callback`	`lambda`	Epoch callback

Returns: train/result · Runs on: any

Example

lolo

(train/loop {"layers":[{"type":"dense","units":2}]} [{"state":[0],"action":1,"reward":1}] {"epochs":10})

`train/mse`

Mean squared error loss

lolo

(train/mse predicted target)

Parameters

Name	Type	Description
`predicted`	`array`	Predicted values
`target`	`array`	Target values

Returns: number

Example

lolo

(train/mse [1.0 2.0] [1.1 2.1]) // => 0.01

`train/sampleBatch`

Sample random batch from experience buffer

lolo

(train/sampleBatch buffer batch_size)

Parameters

Name	Type	Description
`buffer`	`array`	Experience buffer
`batch_size`	`number`	Batch size

Returns: array

Example

lolo

(train/sampleBatch [{"state":[0],"reward":1},{"state":[1],"reward":0}] 2)

`train/sgd`

Stochastic gradient descent optimizer step

lolo

(train/sgd gradients learning_rate)

Parameters

Name	Type	Description
`gradients`	`object`	Gradient structure
`learning_rate`	`number`	Learning rate (variadic)

Returns: void · Runs on: any

Example

lolo

(train/sgd {"layers":[{"type":"dense","units":2}]} 0.01 0.9)

`train/step`

Execute single training step (forward, loss, backward, update)

lolo

(train/step model input target optimizer)

Parameters

Name	Type	Description
`model`	`object`	Neural network model
`input`	`any`	Input data
`target`	`any`	Target data
`optimizer`	`object`	Optimizer state

Returns: train/stepResult · Runs on: any

Example

lolo

(train/step {"layers":[{"type":"dense","units":2}]} [0.1,0.2] [1] {"lr":0.01})

`train/validate`

Validate model on test cases, returns pass/fail metrics

lolo

(train/validate model test_data)

Parameters

Name	Type	Description
`model`	`object`	Neural network model
`test_data`	`array`	Test dataset

Returns: train/validationResult

Example

lolo

(train/validate {"layers":[{"type":"dense","units":2}]} [{"input":[0],"expected":[1]}])

Spatial

Host

Orb

The language where the rule is the program.

Learn

Get started Docs Blog

Tools

Playground Standard library Explore Play Downloads Studio

Community

GitHub Discord Almadar