Blob storage export field reference

Use this reference when building a consumer for the blob storage export. Types match JSON and JSONL output. Timestamps use YYYY-MM-DD HH:MM:SS.ffffff in UTC.

Exported files

File	When it is exported
`observations_v2/`	Current enriched export
`scores/`	Every export
`traces/`	Deprecated legacy export
`observations/`	Deprecated legacy export

For configuration and migration steps, see Export to blob storage.

Enriched observations (`observations_v2/`)

Each row represents one observation and includes its trace context. Only the selected field groups appear; core is always included.

Field	Type	Description
`id`	string	Unique observation identifier.
`trace_id`	string	Trace identifier shared by related observations and scores.
`project_id`	string	Langfuse project identifier.
`environment`	string	Environment label.
`type`	string	Observation type: `SPAN`, `GENERATION`, `EVENT`, `AGENT`, `TOOL`, `CHAIN`, `RETRIEVER`, `EVALUATOR`, `EMBEDDING`, or `GUARDRAIL`.
`parent_observation_id`	string	Parent observation identifier; empty for a root observation.
`is_root_observation`	boolean	Whether the observation is a logical root.
`start_time`	string (timestamp)	When the observation started.
`end_time`	string (timestamp) or null	When the observation ended.
`name`	string	User-defined observation name.
`metadata`	object	User-supplied observation metadata.
`level`	string	`DEBUG`, `DEFAULT`, `WARNING`, or `ERROR`.
`status_message`	string	Status or error message.
`version`	string	User-defined version.
`input`	string	Observation input; may contain plain text or JSON.
`output`	string	Observation output; may contain plain text or JSON.
`provided_model_name`	string	Model name supplied by the SDK or user.
`model_parameters`	string	Model parameters encoded as JSON.
`usage_details`	object (string → integer)	Token usage by category, such as `input`, `output`, and `total`.
`cost_details`	object (string → number)	Cost in USD by category.
`completion_start_time`	string (timestamp) or null	When the first streamed token was generated.
`prompt_name`	string	Langfuse prompt name.
`prompt_version`	integer or null	Langfuse prompt version.
`total_cost`	number	Total observation cost in USD; `0` when no cost was recorded.
`latency`	number or null	Duration in seconds.
`time_to_first_token`	number or null	Time to first token in seconds.
`model_id`	string	Matched Langfuse model definition identifier.
`created_at`	string (timestamp)	Row creation time.
`updated_at`	string (timestamp)	Last row update time.
`prompt_id`	string	Langfuse prompt identifier.
`tool_calls`	array of strings	Tool calls encoded as JSON strings.
`tool_call_names`	array of strings	Names of called tools.
`tool_definitions`	object	Tool or function schemas supplied to the model.
`usage_pricing_tier_id`	string or null	Pricing tier identifier used for cost calculation.
`usage_pricing_tier_name`	string or null	Pricing tier name used for cost calculation.
`input_price`	string or null	Matched per-unit input price; omitted from Parquet.
`output_price`	string or null	Matched per-unit output price; omitted from Parquet.
`total_price`	string or null	Matched flat per-call price; omitted from Parquet.
`user_id`	string	End-user identifier from the trace.
`session_id`	string	Session identifier from the trace.
`trace_name`	string	Trace name.
`tags`	array of strings	Trace tags.
`release`	string	Trace release.
`bookmarked`	boolean	Whether the trace is bookmarked.
`public`	boolean	Whether the trace is public.

Integrations created on or after 2026-04-01 export latency and time_to_first_token in seconds. Older integrations export these fields in milliseconds for backward compatibility.

Scores (`scores/`)

Scores are always exported. Their fields are not configurable. Scores with data type NUMERIC, BOOLEAN, CATEGORICAL, or TEXT are included; corrections (data type CORRECTION) are not.

Field	Type	Description
`id`	string	Unique score identifier.
`timestamp`	string (timestamp)	Score creation time.
`project_id`	string	Langfuse project identifier.
`environment`	string	Environment label.
`trace_id`	string or null	Associated trace identifier.
`observation_id`	string or null	Associated observation identifier.
`session_id`	string or null	Associated session identifier.
`dataset_run_id`	string or null	Associated dataset run identifier.
`name`	string	Score name.
`value`	number	Numeric value; `TEXT` scores use `0`.
`source`	string	`API`, `ANNOTATION`, or `EVAL`.
`comment`	string or null	Optional comment or evaluator reasoning.
`data_type`	string	`NUMERIC`, `BOOLEAN`, `CATEGORICAL`, or `TEXT`.
`string_value`	string or null	Category label or text value; null for numeric scores.
`created_at`	string (timestamp)	Row creation time.
`updated_at`	string (timestamp)	Last row update time.

Legacy exports are deprecated. They split trace context into traces/ and observation data into observations/; consumers join them on trace_id. See upgrade a legacy export before changing a consumer.

Traces (`traces/`)

The trace file has a fixed schema; field groups do not apply.

Field	Type	Description
`id`	string	Unique trace identifier.
`timestamp`	string (timestamp)	Trace creation time.
`name`	string	User-defined trace name.
`environment`	string	Environment label.
`project_id`	string	Langfuse project identifier.
`metadata`	object	Trace metadata.
`user_id`	string or null	End-user identifier.
`session_id`	string or null	Session identifier.
`release`	string or null	Application release.
`version`	string or null	User-defined version.
`public`	boolean	Whether the trace is public.
`bookmarked`	boolean	Whether the trace is bookmarked.
`tags`	array of strings	Trace tags.
`input`	string or null	Trace input.
`output`	string or null	Trace output.
`created_at`	string (timestamp)	Row creation time.
`updated_at`	string (timestamp)	Last row update time.

Legacy trace-level input, output, metadata, timestamp, and version do not have direct equivalents in observations_v2/. Observation fields with the same names contain observation-level data instead.

The trace file does not include total_cost, latency, observations, scores, or html_path.

Observations (`observations/`)

Each row represents one observation without its trace context; join trace_id to the traces/ file to add it. Only the selected field groups appear; core is always included.

Field	Type	Description
`id`	string	Unique observation identifier.
`trace_id`	string	Trace identifier shared by related observations and scores.
`project_id`	string	Langfuse project identifier.
`environment`	string	Environment label.
`type`	string	Observation type: `SPAN`, `GENERATION`, `EVENT`, `AGENT`, `TOOL`, `CHAIN`, `RETRIEVER`, `EVALUATOR`, `EMBEDDING`, or `GUARDRAIL`.
`parent_observation_id`	string or null	Parent observation identifier; null for a root observation.
`start_time`	string (timestamp)	When the observation started.
`end_time`	string (timestamp) or null	When the observation ended.
`name`	string	User-defined observation name.
`metadata`	object	User-supplied observation metadata.
`level`	string	`DEBUG`, `DEFAULT`, `WARNING`, or `ERROR`.
`status_message`	string or null	Status or error message.
`version`	string or null	User-defined version.
`input`	string or null	Observation input; may contain plain text or JSON.
`output`	string or null	Observation output; may contain plain text or JSON.
`provided_model_name`	string or null	Model name supplied by the SDK or user.
`model_parameters`	string or null	Model parameters encoded as JSON.
`usage_details`	object (string → integer)	Token usage by category, such as `input`, `output`, and `total`.
`cost_details`	object (string → number)	Cost in USD by category.
`completion_start_time`	string (timestamp) or null	When the first streamed token was generated.
`prompt_name`	string or null	Langfuse prompt name.
`prompt_version`	integer or null	Langfuse prompt version.
`total_cost`	number or null	Total observation cost in USD.
`latency`	number or null	Duration in seconds.
`time_to_first_token`	number or null	Time to first token in seconds.
`model_id`	string or null	Matched Langfuse model definition identifier.
`created_at`	string (timestamp)	Row creation time.
`updated_at`	string (timestamp)	Last row update time.
`prompt_id`	string or null	Langfuse prompt identifier.
`tool_calls`	array of strings	Tool calls encoded as JSON strings.
`tool_call_names`	array of strings	Names of called tools.
`tool_definitions`	object	Tool or function schemas supplied to the model.
`usage_pricing_tier_name`	string or null	Pricing tier name used for cost calculation.
`input_price`	string or null	Matched per-unit input price; omitted from Parquet.
`output_price`	string or null	Matched per-unit output price; omitted from Parquet.
`total_price`	string or null	Matched flat per-call price; omitted from Parquet.

Where a field appears in both observation files it carries the same meaning. Unset values are encoded differently, though: a field that is null above maps to one of three values in observations_v2/ JSON and JSONL output.

Value in `observations_v2/`	Fields
`""` (empty string)	`input`, `model_id`, `model_parameters`, `output`, `parent_observation_id`, `prompt_id`, `prompt_name`, `provided_model_name`, `status_message`, `version`
`null`	`completion_start_time`, `end_time`, `input_price`, `latency`, `output_price`, `prompt_version`, `time_to_first_token`, `total_price`, `usage_pricing_tier_name`
`0`	`total_cost`

Empty strings appear where the v4 events table stores the column as non-nullable. total_cost reads from cost_details['total'], so a 0 there cannot be told apart from a genuine zero cost.

For the field groups that select these columns and how they differ from the enriched export, see what changes in the exported data.

Parquet differences

Parquet uses internal encoding and compression, so gzip does not apply.

Observation files omit input_price, output_price, and total_price. Use cost_details and total_cost for cost data.

Trace and score fields are the same across all formats.

File organization

See process exports for paths, filenames, and the manifest workflow.

Was this page helpful?

On this page