LLM Mesh Core#

This page groups the native LLM Mesh objects and query/response builders around the main DSSLLM entry point.

Core objects#

class dataikuapi.dss.llm.DSSLLM(client, project_key, llm_id)#

A handle to interact with a DSS-managed LLM.

Important

Do not create this class directly, use dataikuapi.dss.project.DSSProject.get_llm() instead.

new_completion()#

Create a new completion query.

Returns:

A handle on the generated completion query.

Return type:

DSSLLMCompletionQuery

new_completions()#

Create a new multi-completion query.

Returns:

A handle on the generated multi-completion query.

Return type:

DSSLLMCompletionsQuery

new_embeddings(text_overflow_mode='FAIL')#

Create a new embedding query.

Parameters:

text_overflow_mode (str) – How to handle longer texts than what the model supports. Either ‘TRUNCATE’ or ‘FAIL’.

Returns:

A handle on the generated embeddings query.

Return type:

DSSLLMEmbeddingsQuery

new_images_generation()#
new_reranking()#

Create a new reranking query.

Returns:

A handle on the generated reranking query.

Return type:

DSSLLMRerankingQuery

as_langchain_llm(**data)#

Create a langchain-compatible LLM object for this LLM.

Returns:

A langchain-compatible LLM object.

Return type:

dataikuapi.dss.langchain.llm.DKULLM

as_langchain_chat_model(**data)#

Create a langchain-compatible chat LLM object for this LLM.

Returns:

A langchain-compatible LLM object.

Return type:

dataikuapi.dss.langchain.llm.DKUChatModel

as_langchain_embeddings(**data)#

Create a langchain-compatible embeddings object for this LLM.

Returns:

A langchain-compatible embeddings object.

Return type:

dataikuapi.dss.langchain.embeddings.DKUEmbeddings

class dataikuapi.dss.llm.DSSLLMListItem(client, project_key, data)#

An item in a list of llms

Important

Do not instantiate this class directly, instead use dataikuapi.dss.project.DSSProject.list_llms().

to_llm()#

Convert the current item.

Returns:

A handle for the llm.

Return type:

dataikuapi.dss.llm.DSSLLM

property id#
Returns:

The id of the llm.

Return type:

string

property type#
Returns:

The type of the LLM

Return type:

string

property description#
Returns:

The description of the LLM

Return type:

string

Text generation#

class dataikuapi.dss.llm.DSSLLMCompletionQuery(llm)#

A handle to interact with a completion query. Completion queries allow you to send a prompt to a DSS-managed LLM and retrieve its response.

Important

Do not create this class directly, use dataikuapi.dss.llm.DSSLLM.new_completion() instead.

property settings#
Returns:

The completion query settings.

Return type:

dict

new_guardrail(type)#

Start adding a guardrail to the request. You need to configure the returned object, and call add() to actually add it

execute()#

Run the completion query and retrieve the LLM response.

Returns:

The LLM response.

Return type:

DSSLLMCompletionResponse

execute_streamed(collect_response=False)#

Run the completion query and retrieve the LLM response as streamed chunks.

Parameters:

collect_response (bool) – If True, the streamed chunks are also aggregated into a consolidated DSSLLMCompletionResponse by the returned iterator.

Returns:

An iterator over the LLM response chunks

Return type:

DSSLLMStreamedCompletionChunks

new_multipart_message(role='user')#

Start adding a multipart-message to the completion query.

Use this to add image parts to the message.

Parameters:

role (str) – The message role. Use system to set the LLM behavior, assistant to store predefined responses, user to provide requests or comments for the LLM to answer to. Defaults to user.

Return type:

DSSLLMCompletionQueryMultipartMessage

new_multipart_tool_output(tool_call_id, role='tool', output='')#

Start adding a multipart tool output to the completion query.

Parameters:

  • tool_call_id (str) – The tool call id, as provided by the LLM in the conversation messages.

  • role (str) – The message role. Defaults to tool.

  • output (str) – The tool’s output. Defaults to an empty string.

Return type:

DSSLLMCompletionQueryMultipartToolOutput

with_context(context)#
with_json_output(schema=None, strict=None, compatible=None)#

Request the model to generate a valid JSON response, for models that support it.

Note that some models may require you to also explicitly request this in the user or system prompt to use this.

Caution

JSON output support is experimental for locally-running Hugging Face models.

Parameters:

  • schema (dict) – (optional) If specified, request the model to produce a JSON response that adheres to the provided schema. Support varies across models/providers.

  • strict (bool) – (optional) If a schema is provided, whether to strictly enforce it. Support varies across models/providers.

  • compatible (bool) – (optional) Allow DSS to modify the schema in order to increase compatibility, depending on known limitations of the model/provider. Defaults to automatic.

with_memory_fragment(memory_fragment)#

Add a memory fragment to the completion query.

Parameters:

memory_fragment (dict) – The memory fragment returned by the model on the previous turn.

with_message(message, role='user')#

Add a message to the completion query.

Parameters:

  • message (str) – The message text.

  • role (str) – The message role. Use system to set the LLM behavior, assistant to store predefined responses, user to provide requests or comments for the LLM to answer to. Defaults to user.

with_structured_output(model_type, strict=None, compatible=None)#

Instruct the model to generate a response as an instance of a specified Pydantic model.

This functionality depends on with_json_output and necessitates that the model supports JSON output with a schema.

Caution

Structured output support is experimental for locally-running Hugging Face models.

Parameters:

  • model_type (pydantic.BaseModel) – A Pydantic model class used for structuring the response.

  • strict (bool) – (optional) see with_json_output()

  • compatible (bool) – (optional) see with_json_output()

with_tool_calls(tool_calls, role='assistant')#

Add tool calls to the completion query.

Caution

Tool calls support is experimental for locally-running Hugging Face models.

Parameters:

  • tool_calls (list[dict]) – Calls to tools that the LLM requested to use.

  • role (str) – The message role. Defaults to assistant.

with_tool_output(tool_output, tool_call_id, role='tool')#

Add a tool message to the completion query.

Parameters:

  • tool_output (str) – The tool output, as a string.

  • tool_call_id (str) – The tool call id, as provided by the LLM in the conversation messages.

  • role (str) – The message role. Defaults to tool.

with_tool_validation_requests(tool_validation_requests)#

Add tool validation requests to the completion query.

Parameters:

tool_validation_requests (list[dict]) – Validation requests for tools that the agent requested to use.

with_tool_validation_response(validation_request_id, validated=True, arguments=None)#

Add a tool validation response to the completion query.

Parameters:

  • validation_request_id (str) – The validation request id, as provided by the agent in the conversation messages.

  • validated (bool) – Whether to validate or reject the tool call.

  • arguments (str) – Arguments to use for the tool call (if different from the validation request).

class dataikuapi.dss.llm.DSSLLMCompletionResponse(raw_resp=None, text=None, finish_reason=None, response_parser=None, trace=None, query=None)#

Response to a completion

property json#
Returns:

LLM response parsed as a JSON object

property parsed#
property success#
Returns:

The outcome of the completion query.

Return type:

bool

property text#
Returns:

The raw text of the LLM response.

Return type:

Union[str, None]

property tool_calls#
Returns:

The tool calls of the LLM response.

Return type:

Union[list, None]

property tool_validation_requests#
Returns:

The tool validation requests of the agent response.

Return type:

Union[list, None]

property memory_fragment#
Returns:

Data generated by the model that must be passed back in the next query.

Return type:

Union[dict, None]

property log_probs#
Returns:

The log probs of the LLM response.

Return type:

Union[list, None]

property context_upsert#
Returns:

The context upsert of the response (only for agents).

Return type:

Union[dict, None]

property trace#
property total_usage#
prepare_followup()#

Prepare a new completion query to follow up on this response, pre-filled with the relevant data from the response.

Returns:

The prepared follow-up completion query.

Return type:

DSSLLMCompletionQuery

class dataikuapi.dss.llm.DSSLLMStreamedCompletionChunks(query, collect_response=False)#

An iterator over the chunks generated by the execution of a streamed completion query. The streamed chunks are of type DSSLLMStreamedCompletionChunk and DSSLLMStreamedCompletionFooter. When collect_response=True, the streamed chunks are aggregated into a consolidated DSSLLMCompletionResponse.

Important

Do not create this class directly, use dataikuapi.dss.llm.DSSLLMCompletionQuery.execute_streamed() instead.

iter_chunks()#
Returns:

An iterator over the LLM response chunks.

Return type:

Iterator[Union[DSSLLMStreamedCompletionChunk, DSSLLMStreamedCompletionFooter]]

property response#
Returns:

The consolidated LLM response obtained by the aggregation of all streamed chunks, if collect_response=True. Available only after all chunks have been collected.

Return type:

DSSLLMCompletionResponse

prepare_followup()#

Prepare a followup completion query from the consolidated response, pre-filled with the relevant data from the response. Available only when collect_response=True, after all chunks have been collected.

Returns:

The prepared follow-up completion query.

Return type:

DSSLLMCompletionQuery

Batch text generation#

class dataikuapi.dss.llm.DSSLLMCompletionsQuery(llm)#

A handle to interact with a multi-completion query. Completion queries allow you to send a prompt to a DSS-managed LLM and retrieve its response.

Important

Do not create this class directly, use dataikuapi.dss.llm.DSSLLM.new_completion() instead.

property settings#
Returns:

The completion query settings.

Return type:

dict

new_completion()#
new_guardrail(type)#

Start adding a guardrail to the request. You need to configure the returned object, and call add() to actually add it

execute()#

Run the completions query and retrieve the LLM response.

Returns:

The LLM response.

Return type:

DSSLLMCompletionsResponse

with_json_output(schema=None, strict=None, compatible=None)#

Request the model to generate a valid JSON response, for models that support it.

Note that some models may require you to also explicitly request this in the user or system prompt to use this.

Caution

JSON output support is experimental for locally-running Hugging Face models.

Parameters:

  • schema (dict) – (optional) If specified, request the model to produce a JSON response that adheres to the provided schema. Support varies across models/providers.

  • strict (bool) – (optional) If a schema is provided, whether to strictly enforce it. Support varies across models/providers.

  • compatible (bool) – (optional) Allow DSS to modify the schema in order to increase compatibility, depending on known limitations of the model/provider. Defaults to automatic.

with_structured_output(model_type, strict=None, compatible=None)#

Instruct the model to generate a response as an instance of a specified Pydantic model.

This functionality depends on with_json_output and necessitates that the model supports JSON output with a schema.

Caution

Structured output support is experimental for locally-running Hugging Face models.

Parameters:

  • model_type (pydantic.BaseModel) – A Pydantic model class used for structuring the response.

  • strict (bool) – (optional) see with_json_output()

  • compatible (bool) – (optional) see with_json_output()

class dataikuapi.dss.llm.DSSLLMCompletionsQuerySingleQuery#
new_multipart_message(role='user')#

Start adding a multipart-message to the completion query.

Use this to add image parts to the message.

Parameters:

role (str) – The message role. Use system to set the LLM behavior, assistant to store predefined responses, user to provide requests or comments for the LLM to answer to. Defaults to user.

Return type:

DSSLLMCompletionQueryMultipartMessage

with_message(message, role='user')#

Add a message to the completion query.

Parameters:

  • message (str) – The message text.

  • role (str) – The message role. Use system to set the LLM behavior, assistant to store predefined responses, user to provide requests or comments for the LLM to answer to. Defaults to user.

with_memory_fragment(memory_fragment)#

Add a memory fragment to the completion query.

Parameters:

memory_fragment (dict) – The memory fragment returned by the model on the previous turn.

with_tool_calls(tool_calls, role='assistant')#

Add tool calls to the completion query.

Caution

Tool calls support is experimental for locally-running Hugging Face models.

Parameters:

  • tool_calls (list[dict]) – Calls to tools that the LLM requested to use.

  • role (str) – The message role. Defaults to assistant.

with_tool_validation_requests(tool_validation_requests)#

Add tool validation requests to the completion query.

Parameters:

tool_validation_requests (list[dict]) – Validation requests for tools that the agent requested to use.

with_tool_validation_response(validation_request_id, validated=True, arguments=None)#

Add a tool validation response to the completion query.

Parameters:

  • validation_request_id (str) – The validation request id, as provided by the agent in the conversation messages.

  • validated (bool) – Whether to validate or reject the tool call.

  • arguments (str) – Arguments to use for the tool call (if different from the validation request).

new_multipart_tool_output(tool_call_id, role='tool', output='')#

Start adding a multipart tool output to the completion query.

Parameters:

  • tool_call_id (str) – The tool call id, as provided by the LLM in the conversation messages.

  • role (str) – The message role. Defaults to tool.

  • output (str) – The tool’s output. Defaults to an empty string.

Return type:

DSSLLMCompletionQueryMultipartToolOutput

with_tool_output(tool_output, tool_call_id, role='tool')#

Add a tool message to the completion query.

Parameters:

  • tool_output (str) – The tool output, as a string.

  • tool_call_id (str) – The tool call id, as provided by the LLM in the conversation messages.

  • role (str) – The message role. Defaults to tool.

with_context(context)#
class dataikuapi.dss.llm.DSSLLMCompletionsResponse(raw_resp, response_parser=None)#

A handle to interact with a multi-completion response.

Important

Do not create this class directly, use dataikuapi.dss.llm.DSSLLMCompletionsQuery.execute() instead.

property responses#

The array of responses

Multipart message and tool payload builders#

class dataikuapi.dss.llm.DSSLLMCompletionQueryMultipartMessage(q, role)#

Important

Do not create this class directly, use dataikuapi.dss.llm.DSSLLMCompletionQuery.new_multipart_message() or dataikuapi.dss.llm.DSSLLMCompletionsQuerySingleQuery.new_multipart_message().

add()#

Add this message to the completion query

class dataikuapi.dss.llm.DSSLLMCompletionQueryMultipartToolOutput(q, tool_call_id, role, output)#

Important

Do not create this class directly, use dataikuapi.dss.llm.DSSLLMCompletionQuery.new_multipart_tool_output() or dataikuapi.dss.llm.DSSLLMCompletionsQuerySingleQuery.new_multipart_tool_output().

add()#

Add this tool output to the completion query

Embeddings#

class dataikuapi.dss.llm.DSSLLMEmbeddingsQuery(llm, text_overflow_mode)#

A handle to interact with an embedding query. Embedding queries allow you to transform text into embedding vectors using a DSS-managed model.

Important

Do not create this class directly, use dataikuapi.dss.llm.DSSLLM.new_embeddings() instead.

add_text(text)#

Add text to the embedding query.

Parameters:

text (str) – Text to add to the query.

add_image(image, text=None)#

Add an image to the embedding query.

Parameters:

  • image – Image content as bytes or str (base64)

  • text – Optional text (requires a multimodal model)

new_guardrail(type)#

Start adding a guardrail to the request. You need to configure the returned object, and call add() to actually add it

execute()#

Run the embedding query.

Returns:

The results of the embedding query.

Return type:

DSSLLMEmbeddingsResponse

class dataikuapi.dss.llm.DSSLLMEmbeddingsResponse(raw_resp)#

A handle to interact with an embedding query result.

Important

Do not create this class directly, use dataikuapi.dss.llm.DSSLLMEmbeddingsQuery.execute() instead.

get_embeddings()#

Retrieve vectors resulting from the embeddings query.

Returns:

A list of lists containing all embedding vectors.

Return type:

list

Image generation#

class dataikuapi.dss.llm.DSSLLMImageGenerationQuery(llm)#

A handle to interact with an image generation query.

Important

Do not create this class directly, use dataikuapi.dss.llm.DSSLLM.new_images_generation() instead.

with_prompt(prompt, weight=None)#

Add a prompt to the image generation query.

Parameters:

  • prompt (str) – The prompt text.

  • weight (float) – Optional weight between 0 and 1 for the prompt.

with_negative_prompt(prompt, weight=None)#

Add a negative prompt to the image generation query.

Parameters:

  • prompt (str) – The prompt text.

  • weight (float) – Optional weight between 0 and 1 for the negative prompt.

with_original_image(image, mode=None, weight=None)#

Add an image to the generation query.

To edit specific pixels of the original image. A mask can be applied by calling with_mask():

>>> query.with_original_image(image, mode="INPAINTING") # replace the pixels using a mask

To edit an image:

>>> query.with_original_image(image, mode="MASK_FREE") # edit the original image according to the prompt

>>> query.with_original_image(image, mode="VARY") # generates a variation of the original image
Parameters:

  • image (Union[str, bytes]) – The original image as str in base 64 or bytes.

  • mode (str) – The edition mode. Modes support varies across models/providers.

  • weight (float) – The original image weight between 0 and 1.

with_mask(mode, image=None)#

Add a mask for edition to the generation query. Call this method alongside with_original_image().

To edit parts of the image using a black mask (replace the black pixels):

>>> query.with_mask("MASK_IMAGE_BLACK", image=black_mask)

To edit parts of the image that are transparent (replace the transparent pixels):

>>> query.with_mask("ORIGINAL_IMAGE_ALPHA")
Parameters:

  • mode (str) – The mask mode. Modes support varies across models/providers.

  • image (Union[str, bytes]) – The mask image to apply to the image edition. As str in base 64 or bytes.

new_guardrail(type)#

Start adding a guardrail to the request. You need to configure the returned object, and call add() to actually add it

property height#
Returns:

The generated image height in pixels.

Return type:

Optional[int]

property width#
Returns:

The generated image width in pixels.

Return type:

Optional[int]

property fidelity#
Returns:

From 0.0 to 1.0, how strongly to adhere to prompt.

Return type:

Optional[float]

property quality#
Returns:

Quality of the image to generate. Valid values depend on the targeted model.

Return type:

Optional[str]

property seed#
Returns:

Seed of the image to generate, gives deterministic results when set.

Return type:

Optional[int]

property style#
Returns:

Style of the image to generate. Valid values depend on the targeted model.

Return type:

Optional[str]

property images_to_generate#
Returns:

Number of images to generate per query. Valid values depend on the targeted model.

Return type:

Optional[int]

property aspect_ratio#
Returns:

The width/height aspect ratio or None if either is not set.

Return type:

Optional[float]

execute()#

Executes the image generation

Return type:

DSSLLMImageGenerationResponse

class dataikuapi.dss.llm.DSSLLMImageGenerationResponse(raw_resp)#

A handle to interact with an image generation response.

Important

Do not create this class directly, use dataikuapi.dss.llm.DSSLLMImageGenerationQuery.execute() instead.

property success#
Returns:

The outcome of the image generation query.

Return type:

bool

first_image(as_type='bytes')#
Parameters:

as_type (str) – The type of image to return, ‘bytes’ for bytes otherwise ‘str’ for base 64 str.

Returns:

The first generated image as bytes or str depending on the as_type parameter.

Return type:

Union[bytes,str]

get_images(as_type='bytes')#
Parameters:

as_type (str) – The type of images to return, ‘bytes’ for bytes otherwise ‘str’ for base 64 str.

Returns:

The generated images as bytes or str depending on the as_type parameter.

Return type:

Union[List[bytes], List[str]]

property images#
Returns:

The generated images in bytes format.

Return type:

List[bytes]

property trace#
property total_usage#

Reranking#

class dataikuapi.dss.llm.DSSLLMRerankingQuery(llm)#

A handle to interact with a reranking query. Reranking queries allow you to send a text query and a list of documents to a DSS-managed ranking model and retrieve the documents ranked according to their relevance to the query.

Important

Do not create this class directly, use dataikuapi.dss.llm.DSSLLM.new_reranking() instead.

with_query(text)#

Sets the reranking text query.

Parameters:

text (str) – The reranking text query.

with_document(text)#

Adds a text document to the list of documents to be reranked.

Parameters:

text (str) – The text document to be reranked.

execute()#

Run the reranking query and retrieve the LLM response.

Returns:

The LLM response.

Return type:

DSSLLMRerankingResponse

class dataikuapi.dss.llm.DSSLLMRerankingResponse(raw_resp)#

A handle to interact with a ranking query result.

Important

Do not create this class directly, use dataikuapi.dss.llm.DSSLLMRerankingQuery.execute() instead.

property success#
Returns:

The outcome of the reranking query.

Return type:

bool

property error_message#
Returns:

The error message if the reranking query failed, None otherwise.

Return type:

Union[str, None]

property documents#
Returns:

The array of reranked documents.

Return type:

list of DSSLLMRerankingResponse.RankedDocument

property trace#
Returns:

The trace of the reranking query if available, None otherwise.

Return type:

Union[dict, None]

class RankedDocument(raw_doc)#
property index#
Returns:

The index of the document in the original request.

Return type:

int

property relevance_score#
Returns:

The relevance score assigned to the document by the ranking model.

Return type:

float