feat(api): add chunking mode to enrichment

Author: stainless-app[bot]

.stats.yml

@@ -1,4 +1,4 @@
 configured_endpoints: 5
-openapi_spec_url: https://storage.googleapis.com/stainless-sdk-openapi-specs/isaacus%2Fisaacus-7eb7a36b7ad1dc6d3a183088ede90dfbef328d230641e38de953ccc8dd81d137.yml
-openapi_spec_hash: 9ef218f1e50c15715e174c95cb6d153c
+openapi_spec_url: https://storage.googleapis.com/stainless-sdk-openapi-specs/isaacus%2Fisaacus-46de7b353c33e2b6c139e564caeec9e0462ad714121690f65167a7943e325000.yml
+openapi_spec_hash: 6527287f9709a8741c9cc5b4181d7bb1
 config_hash: 9040e7359f066240ad536041fb2c5185

src/isaacus/resources/enrichments.py

@@ -49,7 +49,7 @@ def create(
         *,
         model: Literal["kanon-2-enricher"],
         texts: Union[SequenceNotStr[str], str],
-        overflow_strategy: Optional[Literal["auto", "drop_end"]] | Omit = omit,
+        overflow_strategy: Optional[Literal["auto", "drop_end", "chunk"]] | Omit = omit,
         # Use the following arguments if you need to pass additional parameters to the API that aren't available via kwargs.
         # The extra values given here take precedence over values defined on the client or passed to this method.
         extra_headers: Headers | None = None,
@@ -71,9 +71,15 @@ def create(
 
           overflow_strategy: The strategy for handling content exceeding the model's maximum input length.
 
-              `auto` currently behaves the same as `drop_end`, dropping excess tokens from the
-              end of input. In the future, `auto` may implement more sophisticated strategies
-              such as chunking and context-aware stitching.
+              `auto`, which is the default and recommended setting, currently behaves the same
+              as `chunk`, which intelligently breaks the input up into smaller chunks and then
+              stitches the results back together into a single prediction. In the future
+              `auto` may implement even more sophisticated strategies for handling long
+              contexts such as leveraging chunk overlap and/or a specialized stitching model.
+
+              `chunk` breaks the input up into smaller chunks that fit within the model's
+              context window and then intelligently merges the results into a single
+              prediction at the cost of a minor accuracy drop.
 
               `drop_end` drops tokens from the end of input exceeding the model's maximum
               input length.
@@ -131,7 +137,7 @@ async def create(
         *,
         model: Literal["kanon-2-enricher"],
         texts: Union[SequenceNotStr[str], str],
-        overflow_strategy: Optional[Literal["auto", "drop_end"]] | Omit = omit,
+        overflow_strategy: Optional[Literal["auto", "drop_end", "chunk"]] | Omit = omit,
         # Use the following arguments if you need to pass additional parameters to the API that aren't available via kwargs.
         # The extra values given here take precedence over values defined on the client or passed to this method.
         extra_headers: Headers | None = None,
@@ -153,9 +159,15 @@ async def create(
 
           overflow_strategy: The strategy for handling content exceeding the model's maximum input length.
 
-              `auto` currently behaves the same as `drop_end`, dropping excess tokens from the
-              end of input. In the future, `auto` may implement more sophisticated strategies
-              such as chunking and context-aware stitching.
+              `auto`, which is the default and recommended setting, currently behaves the same
+              as `chunk`, which intelligently breaks the input up into smaller chunks and then
+              stitches the results back together into a single prediction. In the future
+              `auto` may implement even more sophisticated strategies for handling long
+              contexts such as leveraging chunk overlap and/or a specialized stitching model.
+
+              `chunk` breaks the input up into smaller chunks that fit within the model's
+              context window and then intelligently merges the results into a single
+              prediction at the cost of a minor accuracy drop.
 
               `drop_end` drops tokens from the end of input exceeding the model's maximum
               input length.

src/isaacus/types/enrichment_create_params.py

@@ -25,12 +25,18 @@ class EnrichmentCreateParams(TypedDict, total=False):
     No more than 8 texts can be enriched in a single request.
     """
 
-    overflow_strategy: Optional[Literal["auto", "drop_end"]]
+    overflow_strategy: Optional[Literal["auto", "drop_end", "chunk"]]
     """The strategy for handling content exceeding the model's maximum input length.
 
-    `auto` currently behaves the same as `drop_end`, dropping excess tokens from the
-    end of input. In the future, `auto` may implement more sophisticated strategies
-    such as chunking and context-aware stitching.
+    `auto`, which is the default and recommended setting, currently behaves the same
+    as `chunk`, which intelligently breaks the input up into smaller chunks and then
+    stitches the results back together into a single prediction. In the future
+    `auto` may implement even more sophisticated strategies for handling long
+    contexts such as leveraging chunk overlap and/or a specialized stitching model.
+
+    `chunk` breaks the input up into smaller chunks that fit within the model's
+    context window and then intelligently merges the results into a single
+    prediction at the cost of a minor accuracy drop.
 
     `drop_end` drops tokens from the end of input exceeding the model's maximum
     input length.