ChatLLMTrainingConfig

Training config for the CHAT_LLM problem type

KEY	TYPE	Description
ENABLE_LLM_REWRITE	bool	If enabled, an LLM will rewrite the RAG queries sent to document retriever. Disabled by default.
ENABLE_WEB_SEARCH	bool	Allow the LLM to use Web Search Engines to retrieve information for better results.
DATABASE_CONNECTOR_ID	str	Database connector ID to use for connecting external database that gives access to structured data to the LLM.
INCLUDE_GENERAL_KNOWLEDGE	bool	Allow the LLM to rely not just on RAG search results, but to fall back on general knowledge. Disabled by default.
BEHAVIOR_INSTRUCTIONS	str	Customize the overall behaviour of the model. This controls things like - when to execute code (if enabled), write sql query, search web (if enabled), etc.
COLUMN_FILTERING_INSTRUCTIONS	str	Instructions for a LLM call to automatically generate filter expressions on document metadata to retrieve relevant documents for the conversation.
DATABASE_CONNECTOR_IDS	List[str]	List of database connector IDs to use for connecting external databases that give access to structured data to the LLM.
LOOKUP_REWRITE_INSTRUCTIONS	None	None
ENABLE_INLINE_SOURCE_CITATIONS	bool	Enable inline citations of the sources in the response.
MASK_PII	bool	Mask PII in the prompts and uploaded documents before sending it to the LLM. Only available for Enterprise users and will cause validation errors if set to True for ChatLLM Teams users.
JSON_RESPONSE_INSTRUCTIONS	str	Instructions to be followed while generating the json_response if `response_format` is set to "JSON". This can include the schema information if the schema is dynamic and its keys cannot be pre-determined.
DATABASE_CONNECTOR_TABLES	List[str]	List of tables to use from the database connector for the ChatLLM.
HIDE_SQL_AND_CODE	bool	When running data queries, this will hide the generated SQL and Code in the response.
DATA_PROMPT_CONTEXT	str	Prompt context for the data feature group IDs.
NUM_COMPLETION_TOKENS	int	Default for maximum number of tokens for chat answers. Reducing this will get faster responses which are more succinct.
ENABLE_CODE_EXECUTION	bool	Enable python code execution in the ChatLLM. This equips the LLM with a python kernel in which all its code is executed.
BUILTIN_TOOLS	List[SystemConnectorTool]	List of builtin system connector tools to use in the ChatLLM. Using builtin tools does not require enabling tool bar (enable_tool_bar flag).
SEARCH_SCORE_CUTOFF	float	Minimum search score to consider a document as a valid search result.
ENABLE_TOOL_BAR	bool	Enable the tool bar in Enterprise ChatLLM to provide additional functionalities like tool_use, web_search, image_gen, etc. Enabling this requires enable_web_search to be enabled.
UNKNOWN_ANSWER_PHRASE	str	Fallback response when the LLM can't find an answer.
AGENTIC_LOOP_MODE	bool	Enables use of agentic loop that uses a series of tool calls when needed to respond. If set to False, the agentic loop will not be used. If not set or set to Auto, the agentic loop will be automatically used based on certain conditions like presence of tools in the model.
CONFIG_CONNECTORS	List[str]	List of names of config connectors to use in the ChatLLM. This should not be used with document_retrievers.
RETRIEVAL_COLUMNS	list	Include the metadata column values in the retrieved search results.
TEMPERATURE	float	The generative LLM temperature.
DATA_PROMPT_COLUMN_CONTEXT	Dict[str, str]	Dict of 'table_name.column_name' and 'column_context' pairs to provide column context for some selected columns in the selected structured data table. This replaces the default auto-generated information about the column data.
DISABLE_DATA_SUMMARIZATION	bool	After executing a query summarize the reponse and reply back with only the table and query run.
DATA_FEATURE_GROUP_IDS	None	(List[str]): List of feature group IDs to use to possibly query for the ChatLLM. The created ChatLLM is commonly referred to as DataLLM.
DISABLE_DATA_FETCH_FOR_TRAINING	bool	Train using only table and column schema metadata without fetching sample data. This speeds up training but may result in less context for the model.
INCLUDE_BM25_RETRIEVAL	bool	Combine BM25 search score with vector search using reciprocal rank fusion.
FILTER_COLUMNS	list	Allow users to filter the document retrievers on these metadata columns.
DATA_PROMPT_TABLE_CONTEXT	Dict[str, str]	Dict of table name and table context pairs to provide table wise context for each structured data table.
JSON_RESPONSE_SCHEMA	str	Specifies the JSON schema that the model should adhere to if `response_format` is set to "JSON". This should be a json-formatted string where each field of the expected schema is mapped to a dictionary containing the fields 'type', 'required' and 'description'. For example - '{"sample_field": {"type": "integer", "required": true, "description": "Sample Field"}}'
MCP_SERVERS	List[str]	List of names of MCP servers to use in the ChatLLM. This should not be used with document_retrievers.
RESPONSE_INSTRUCTIONS	str	Customized instructions for how the model should respond inlcuding the format, persona and tone of the answers.
DOCUMENT_RETRIEVERS	List[str]	List of names or IDs of document retrievers to use as vector stores of information for RAG responses.
KEYWORD_REQUIREMENT_INSTRUCTIONS	str	Instructions for a LLM call to automatically generate keyword requirements to retrieve relevant documents for the conversation.
QUERY_REWRITE_INSTRUCTIONS	str	Special instructions for the LLM which rewrites the RAG query.
ENABLE_RESPONSE_CACHING	bool	Enable caching of LLM responses to speed up response times and improve reproducibility.
METADATA_COLUMNS	None	None
RESPONSE_FORMAT	None	(str): When set to 'JSON', the LLM will generate a JSON formatted string.
DATA_COLUMNS_TO_IGNORE	List[str]	Columns to ignore while encoding information about structured data tables in context for the LLM. A list of strings of format "."
MAX_SEARCH_RESULTS	int	Maximum number of search results in the retrieval augmentation step. If we know that the questions are likely to have snippets which are easily matched in the documents, then a lower number will help with accuracy.