AI Core Components¶

Bases: OpenAI

A LLM implementation that calls the Anaconda AI Navigator API.

Bases: Llm

A LLM implementation that calls Anthropic models such as Claude.

Bases: MistralAI

A LLM implementation that calls Mistral AI models on Azure.

Bases: Llm, AzureOpenAIMixin

A LLM implementation that uses the Azure OpenAI integration. Inherits from AzureOpenAIMixin which extends OpenAIMixin, so it has access to all OpenAI functionality plus Azure-specific configuration.

Bases: OpenAIMixin

Mixin class for Azure OpenAI functionality that extends OpenAI functionality with Azure-specific configuration.

Bases: BaseModel

Bases: Llm

A LLM implementation that calls Google's Gemini models.

Bases: BaseModel

Bases: Parameterized

Bases: Llm

A LLM implementation using LiteLLM that supports multiple providers through a unified interface.

LiteLLM allows you to call 100+ LLMs using the same OpenAI-compatible input/output format, including providers like OpenAI, Anthropic, Cohere, Hugging Face, Azure, Vertex AI, and more.

Bases: Llm, LlamaCppMixin

A LLM implementation using Llama.cpp Python wrapper together with huggingface_hub to fetch the models.

Bases: ServiceMixin

Mixin class for llama-cpp-python functionality that can be shared between LLM implementations and embedding classes.

Bases: Parameterized

Baseclass for LLM implementations.

An LLM implementation wraps a local or cloud based LLM provider with instructor to enable support for correctly validating Pydantic models.

Bases: TypedDict

Bases: BaseModel

Bases: Llm

A LLM implementation that calls Mistral AI.

Bases: OpenAI

An LLM implementation using the Ollama cloud.

Bases: Llm, OpenAIMixin

An LLM implementation using the OpenAI cloud.

Bases: ServiceMixin

Mixin class for OpenAI functionality that can be shared between LLM implementations and embedding classes.

Bases: BaseModel

Bases: Llm

Bases: Embeddings, AzureOpenAIMixin

AzureOpenAIEmbeddings is an embeddings class that uses the Azure OpenAI API to generate embeddings. Inherits from AzureOpenAIMixin which extends OpenAIMixin, so it has access to all OpenAI functionality plus Azure-specific configuration.

:Example:

embeddings = AzureOpenAIEmbeddings() await embeddings.embed(["Hello, world!", "Goodbye, world!"])

Bases: OpenAIMixin

Mixin class for Azure OpenAI functionality that extends OpenAI functionality with Azure-specific configuration.

Bases: Parameterized

Bases: Embeddings

HuggingFaceEmbeddings is an embeddings class that uses sentence-transformers to generate embeddings from Hugging Face models.

:Example:

embeddings = HuggingFaceEmbeddings() await embeddings.embed(["Hello, world!", "Goodbye, world!"])

Bases: Embeddings, LlamaCppMixin

LlamaCppEmbeddings is an embeddings class that uses the llama-cpp-python library to generate embeddings from GGUF models.

:Example:

embeddings = LlamaCppEmbeddings( ... model_kwargs={ ... "repo_id": "Qwen/Qwen3-Embedding-4B-GGUF", ... "filename": "Qwen3-Embedding-4B-Q4_K_M.gguf", ... "n_ctx": 512, ... "n_batch": 64 ... } ... ) await embeddings.embed(["Hello, world!", "Goodbye, world!"])

Bases: ServiceMixin

Mixin class for llama-cpp-python functionality that can be shared between LLM implementations and embedding classes.

Bases: Embeddings

NumpyEmbeddings is a simple embeddings class that uses a hash function to map n-grams to the vocabulary.

Note that the default hash function is not stable across different Python sessions. If you need a stable hash function, you can set the hash_func, e.g. using murmurhash from the mmh3 package.

:Example:

embeddings = NumpyEmbeddings() await embeddings.embed(["Hello, world!", "Goodbye, world!"])

Bases: Embeddings, OpenAIMixin

OpenAIEmbeddings is an embeddings class that uses the OpenAI API to generate embeddings.

:Example:

embeddings = OpenAIEmbeddings() await embeddings.embed(["Hello, world!", "Goodbye, world!"])

Bases: ServiceMixin

Mixin class for OpenAI functionality that can be shared between LLM implementations and embedding classes.

Bases: VectorStore

Vector store implementation using DuckDB for persistent storage.

:Example:

.. code-block:: python

from lumen.ai.vector_store import DuckDBStore

vector_store = DuckDBStore(uri=':memory:)
vector_store.add_file('https://lumen.holoviz.org')
vector_store.query('LLM', threshold=0.1)

Use upsert to avoid adding duplicate content:

.. code-block:: python

from lumen.ai.vector_store import DuckDBStore

vector_store = DuckDBStore(uri=':memory:)
vector_store.upsert([{'text': 'Hello!', 'metadata': {'source': 'greeting'}}])
# Won't add duplicate if content is similar and metadata matches
vector_store.upsert([{'text': 'Hello!', 'metadata': {'source': 'greeting'}}])

Bases: Parameterized

Bases: Parameterized

Mixin for classes that use prompts with LLMs. Provides parameters and methods for prompt templating and LLM interactions.

Bases: Embeddings

NumpyEmbeddings is a simple embeddings class that uses a hash function to map n-grams to the vocabulary.

Note that the default hash function is not stable across different Python sessions. If you need a stable hash function, you can set the hash_func, e.g. using murmurhash from the mmh3 package.

:Example:

embeddings = NumpyEmbeddings() await embeddings.embed(["Hello, world!", "Goodbye, world!"])

Bases: VectorStore

Vector store implementation using NumPy for in-memory storage.

:Example:

.. code-block:: python

from lumen.ai.vector_store import NumpyVectorStore

vector_store = NumpyVectorStore()
vector_store.add_file('https://lumen.holoviz.org')
vector_store.query('LLM', threshold=0.1)

Use upsert to avoid adding duplicate content:

.. code-block:: python

from lumen.ai.vector_store import NumpyVectorStore

vector_store = NumpyVectorStore()
vector_store.upsert([{'text': 'Hello!', 'metadata': {'source': 'greeting'}}])
# Won't add duplicate if content is similar and metadata matches
vector_store.upsert([{'text': 'Hello!', 'metadata': {'source': 'greeting'}}])

Bases: LLMUser

Abstract base class for a vector store.

Bases: BaseModel

Bases: NodeVisitor

Bases: BaseModel

Bases: JSComponent

A component wrapper for the HTML details element with proper content rendering.

This component provides an expandable details/summary element that can be used to hide/show content. It supports any Panel component as content, including Markdown panes for rendering markdown content.

Bases: BaseModel

Bases: Exception

Raise to indicate missing context for a query.

Bases: Viewer, Component

Pipeline encapsulates filters and transformations applied to a :class:lumen.sources.base.Source table.

A Pipeline ingests data from a :class:lumen.sources.base.Source table or another Pipeline applying the declared :class:lumen.filters.base.Filter, :class:lumen.transforms.base.Transform and :class:lumen.transforms.sql.SQLTransform definitions. It can be used to drive one or more visual outputs or leveraged as a standalone component to encapsulate multiple data processing steps.

Bases: BaseModel

Bases: Exception

Raised when the maximum number of retries is exceeded.

Bases: MultiTypeComponent

Source components provide allow querying all kinds of data.

A Source can return one or more tables queried using the .get_tables method, a description of the data returned by each table in the form of a JSON schema accessible via the .get_schema method and lastly a .get method that allows filtering the data.

The Source base class also implements both in-memory and disk caching which can be enabled if a cache_dir is provided. Data cached to disk is stored as parquet files.