November 27, 2025 | Caesar | API Documentation

Renamed compute_units to reasoning_loops

The compute_units parameter has been renamed to reasoning_loops to better reflect what the parameter controls. This is a breaking change for request payloads.

Before (2025-08-08)

1 {
2   "query": "What is the current price of Bitcoin?",
3   "compute_units": 3
4 }

After (2025-11-27)

1 {
2   "query": "What is the current price of Bitcoin?",
3   "reasoning_loops": 3
4 }

New request parameters

Five new parameters give you finer control over research behavior:

Parameter	Type	Default	Description
`collection_ids`	array	-	IDs of collections to search across
`source_timeout`	integer	60	Maximum seconds to wait for each source
`reasoning_mode`	boolean	true	Enable multi-step reasoning
`allow_early_exit`	boolean	true	Allow completion before max loops if sufficient information gathered
`exclude_social`	boolean	false	Exclude social media sources from research

Collection-based research

You can now associate research requests with file collections using the collection_ids parameter. The research pipeline will search across all files within the specified collections.

1 {
2   "query": "Summarize the key findings from my research papers",
3   "collection_ids": ["a1b2c3d4-5678-90ab-cdef-1234567890ab"],
4   "reasoning_loops": 3
5 }

Combine collection_ids with files to include both collection-based and individual file sources in a single request.

New response fields

Two new fields provide better visibility into job execution:

Field	Type	Description
`reasoning_loops_consumed`	integer	Actual loops used
`running_time`	float	Total execution time in seconds

Removed fields

The score field has been removed from result items. Source relevance is now determined internally and reflected in result ordering.

Migration guide

Replace compute_units with reasoning_loops in all requests
Update response parsing to use reasoning_loops_consumed instead of compute_units_consumed
Remove any logic that depends on the score field in results