Skip to main content
A remote cache stores model responses on Expected Parrot so matching calls can reuse them. It is separate from a local Cache, a saved Results object, and the service that executes a job. Use it to reuse responses across sessions without carrying a local cache file between environments. If your aim is simply to reopen completed data for analysis, retrieve the Results object; use the cache when running a job that may contain matching calls. A matching question alone is insufficient for a cache hit. Model configuration, rendered system and user prompts, and iteration information matter. Changing a scenario, agent instruction, or model parameter can change the call even if the question object is unchanged.

​
Control caching in Python

Remote caching follows your Expected Parrot account settings unless disabled for the run. With authentication configured, inspect the setting in Python:
To run a prepared job with an explicit local cache and remote caching disabled:
disable_remote_inference=True runs orchestration locally; a provider model still makes API calls and can incur costs. disable_remote_cache=True controls hosted cache access separately. Omitting it allows the account setting to apply. A cache_used value records response reuse; it does not identify whether the entry originally came from a local or remote cache.

​
Request fresh responses

To request new model responses instead of reusing cached responses, use fresh=True with hosted execution:
For local orchestration, start with a new Cache() and set disable_remote_cache=True, as in the example above. In this version, fresh=True alone does not bypass entries already in a supplied local cache.

​
Download cache entries for a job

Use the remote job UUID from Background jobs, rather than a Results object UUID. These calls require Expected Parrot access:
The return value from remote_cache_get() is a list of CacheEntry objects, not a Results object or a Cache. Constructing the Cache as above lets you inspect it, save it, or pass it as cache=downloaded_cache to a later run. Only matching model calls can reuse those entries. Save the full Results separately to retain the survey, answers, and study metadata.

​
Download entries by cache key

When you already know the keys, retrieve a specific set instead of all entries for a job:
select_keys must be a nonempty list. To merge retrieved entries into an existing local cache, use cache.add_from_dict({entry.key: entry for entry in entries}). Conflicting entries with the same key raise an error. Remote availability and account access still apply; do not assume every key can be retrieved.

​
CLI alternatives

ep settings shows account settings. --fresh requests new responses for a run. These CLI commands complement the Python retrieval and cache-management methods above. See Jobs execution options for controlling caching during a run.