Add strategy and strategy_prompt params to dedupe

docs/reference/DEDUPE.md

@@ -35,9 +35,60 @@ result = await dedupe(
 print(result.data.head())
 ```
 
+## Strategies
+
+Control what happens after clusters are identified using the `strategy` parameter:
+
+### `select` (default)
+
+Picks the best representative from each cluster. Three columns are added:
+
+- `equivalence_class_id` — rows with the same ID are duplicates of each other
+- `equivalence_class_name` — human-readable label for the cluster
+- `selected` — True for the canonical record in each cluster
+
+```python
+result = await dedupe(
+    input=crm_data,
+    equivalence_relation="Same legal entity",
+    strategy="select",
+    strategy_prompt="Prefer the record with the most complete contact information",
+)
+deduped = result.data[result.data["selected"] == True]
+```
+
+### `identify`
+
+Cluster only — no selection or combining. Useful when you want to review clusters before deciding what to do.
+
+- `equivalence_class_id` — rows with the same ID are duplicates of each other
+- `equivalence_class_name` — human-readable label for the cluster
+
+```python
+result = await dedupe(
+    input=crm_data,
+    equivalence_relation="Same legal entity",
+    strategy="identify",
+)
+```
+
+### `combine`
+
+Synthesizes a single combined row per cluster, merging the best information from all duplicates. Original rows are marked `selected=False`, and new combined rows are added with `selected=True`.
+
+```python
+result = await dedupe(
+    input=crm_data,
+    equivalence_relation="Same legal entity",
+    strategy="combine",
+    strategy_prompt="For each field, keep the most recent and complete value",
+)
+combined = result.data[result.data["selected"] == True]
+```
+
 ## What you get back
 
-Three columns added to your data:
+Three columns added to your data (when using `select` or `combine` strategy):
 
 - `equivalence_class_id` — rows with the same ID are duplicates of each other
 - `equivalence_class_name` — human-readable label for the cluster ("Alexandra Butoi", "Naomi Saphra", etc.)
@@ -73,6 +124,8 @@ Output (selected rows only):
 |------|------|-------------|
 | `input` | DataFrame | Data with potential duplicates |
 | `equivalence_relation` | str | What makes two rows duplicates |
+| `strategy` | str | `"identify"`, `"select"` (default), or `"combine"` |
+| `strategy_prompt` | str | Optional instructions for selection or combining |
 | `session` | Session | Optional, auto-created if omitted |
 
 ## Performance

src/everyrow/generated/models/__init__.py

@@ -3,7 +3,9 @@
 from .agent_map_operation import AgentMapOperation
 from .agent_map_operation_input_type_1_item import AgentMapOperationInputType1Item
 from .agent_map_operation_input_type_2 import AgentMapOperationInputType2
-from .agent_map_operation_response_schema_type_0 import AgentMapOperationResponseSchemaType0
+from .agent_map_operation_response_schema_type_0 import (
+    AgentMapOperationResponseSchemaType0,
+)
 from .billing_response import BillingResponse
 from .create_artifact_request import CreateArtifactRequest
 from .create_artifact_request_data_type_0_item import CreateArtifactRequestDataType0Item
@@ -13,6 +15,7 @@
 from .dedupe_operation import DedupeOperation
 from .dedupe_operation_input_type_1_item import DedupeOperationInputType1Item
 from .dedupe_operation_input_type_2 import DedupeOperationInputType2
+from .dedupe_operation_strategy import DedupeOperationStrategy
 from .error_response import ErrorResponse
 from .error_response_details_type_0 import ErrorResponseDetailsType0
 from .http_validation_error import HTTPValidationError
@@ -39,7 +42,9 @@
 from .single_agent_operation import SingleAgentOperation
 from .single_agent_operation_input_type_1_item import SingleAgentOperationInputType1Item
 from .single_agent_operation_input_type_2 import SingleAgentOperationInputType2
-from .single_agent_operation_response_schema_type_0 import SingleAgentOperationResponseSchemaType0
+from .single_agent_operation_response_schema_type_0 import (
+    SingleAgentOperationResponseSchemaType0,
+)
 from .task_result_response import TaskResultResponse
 from .task_result_response_data_type_0_item import TaskResultResponseDataType0Item
 from .task_result_response_data_type_1 import TaskResultResponseDataType1
@@ -61,6 +66,7 @@
     "DedupeOperation",
     "DedupeOperationInputType1Item",
     "DedupeOperationInputType2",
+    "DedupeOperationStrategy",
     "ErrorResponse",
     "ErrorResponseDetailsType0",
     "HTTPValidationError",

src/everyrow/generated/models/dedupe_operation.py

@@ -7,6 +7,7 @@
 from attrs import define as _attrs_define
 from attrs import field as _attrs_field
 
+from ..models.dedupe_operation_strategy import DedupeOperationStrategy
 from ..types import UNSET, Unset
 
 if TYPE_CHECKING:
@@ -26,11 +27,18 @@ class DedupeOperation:
             list of JSON objects
         equivalence_relation (str): Description of what makes two rows equivalent/duplicates
         session_id (None | Unset | UUID): Session ID. If not provided, a new session is auto-created for this task.
+        strategy (DedupeOperationStrategy | Unset): Controls what happens after duplicate clusters are identified.
+            IDENTIFY = cluster only (no selection/combining), SELECT = pick best representative per cluster (default),
+            COMBINE = synthesize a merged row per cluster from all duplicates. Default: DedupeOperationStrategy.SELECT.
+        strategy_prompt (None | str | Unset): Optional natural-language instructions guiding how the LLM selects or
+            combines rows (only used with SELECT and COMBINE strategies). Example: "Prefer the most complete record".
     """
 
     input_: DedupeOperationInputType2 | list[DedupeOperationInputType1Item] | UUID
     equivalence_relation: str
     session_id: None | Unset | UUID = UNSET
+    strategy: DedupeOperationStrategy | Unset = DedupeOperationStrategy.SELECT
+    strategy_prompt: None | str | Unset = UNSET
     additional_properties: dict[str, Any] = _attrs_field(init=False, factory=dict)
 
     def to_dict(self) -> dict[str, Any]:
@@ -56,6 +64,16 @@ def to_dict(self) -> dict[str, Any]:
         else:
             session_id = self.session_id
 
+        strategy: str | Unset = UNSET
+        if not isinstance(self.strategy, Unset):
+            strategy = self.strategy.value
+
+        strategy_prompt: None | str | Unset
+        if isinstance(self.strategy_prompt, Unset):
+            strategy_prompt = UNSET
+        else:
+            strategy_prompt = self.strategy_prompt
+
         field_dict: dict[str, Any] = {}
         field_dict.update(self.additional_properties)
         field_dict.update(
@@ -66,6 +84,10 @@ def to_dict(self) -> dict[str, Any]:
         )
         if session_id is not UNSET:
             field_dict["session_id"] = session_id
+        if strategy is not UNSET:
+            field_dict["strategy"] = strategy
+        if strategy_prompt is not UNSET:
+            field_dict["strategy_prompt"] = strategy_prompt
 
         return field_dict
 
@@ -125,10 +147,28 @@ def _parse_session_id(data: object) -> None | Unset | UUID:
 
         session_id = _parse_session_id(d.pop("session_id", UNSET))
 
+        _strategy = d.pop("strategy", UNSET)
+        strategy: DedupeOperationStrategy | Unset
+        if isinstance(_strategy, Unset):
+            strategy = UNSET
+        else:
+            strategy = DedupeOperationStrategy(_strategy)
+
+        def _parse_strategy_prompt(data: object) -> None | str | Unset:
+            if data is None:
+                return data
+            if isinstance(data, Unset):
+                return data
+            return cast(None | str | Unset, data)
+
+        strategy_prompt = _parse_strategy_prompt(d.pop("strategy_prompt", UNSET))
+
         dedupe_operation = cls(
             input_=input_,
             equivalence_relation=equivalence_relation,
             session_id=session_id,
+            strategy=strategy,
+            strategy_prompt=strategy_prompt,
         )
 
         dedupe_operation.additional_properties = d

src/everyrow/generated/models/dedupe_operation_strategy.py

@@ -0,0 +1,10 @@
+from enum import Enum
+
+
+class DedupeOperationStrategy(str, Enum):
+    COMBINE = "combine"
+    IDENTIFY = "identify"
+    SELECT = "select"
+
+    def __str__(self) -> str:
+        return str(self.value)

src/everyrow/ops.py

@@ -25,6 +25,7 @@
     CreateArtifactRequestDataType1,
     DedupeOperation,
     DedupeOperationInputType1Item,
+    DedupeOperationStrategy,
     LLMEnumPublic,
     MergeOperation,
     MergeOperationLeftInputType1Item,
@@ -633,16 +634,39 @@ async def dedupe(
     equivalence_relation: str,
     session: Session | None = None,
     input: DataFrame | UUID | TableResult | None = None,
+    strategy: Literal["identify", "select", "combine"] | None = None,
+    strategy_prompt: str | None = None,
 ) -> TableResult:
     """Dedupe a table by removing duplicates using AI.
 
     Args:
-        equivalence_relation: Description of what makes items equivalent
+        equivalence_relation: Natural-language description of what makes two rows
+            equivalent/duplicates. Be as specific as needed — the LLM uses this to
+            reason about equivalence, handling abbreviations, typos, name variations,
+            and entity relationships that string matching cannot capture.
         session: Optional session. If not provided, one will be created automatically.
-        input: The input table (DataFrame, UUID, or TableResult)
+        input: The input table (DataFrame, UUID, or TableResult).
+        strategy: Controls what happens after duplicate clusters are identified.
+            - "identify": Cluster only. Adds `equivalence_class_id` and
+              `equivalence_class_name` columns but does NOT select or remove any rows.
+              Use this when you want to review clusters before deciding what to do.
+            - "select" (default): Picks the best representative row from each cluster.
+              Adds `equivalence_class_id`, `equivalence_class_name`, and `selected`
+              columns. Rows with `selected=True` are the canonical records. To get the
+              deduplicated table: `result.data[result.data["selected"] == True]`.
+            - "combine": Synthesizes a single combined row per cluster by merging the
+              best information from all duplicates. Original rows are kept with
+              `selected=False`, and new combined rows are appended with `selected=True`.
+              Useful when no single row has all the information (e.g., one row has the
+              email, another has the phone number).
+        strategy_prompt: Optional natural-language instructions that guide how the LLM
+            selects or combines rows. Only used with "select" and "combine" strategies.
+            Examples: "Prefer the record with the most complete contact information",
+            "For each field, keep the most recent and complete value",
+            "Prefer records from the CRM system over spreadsheet imports".
 
     Returns:
-        TableResult containing the deduped table
+        TableResult containing the deduped table with cluster metadata columns.
     """
     if input is None:
         raise EveryrowError("input is required for dedupe")
@@ -652,6 +676,8 @@ async def dedupe(
                 session=internal_session,
                 input=input,
                 equivalence_relation=equivalence_relation,
+                strategy=strategy,
+                strategy_prompt=strategy_prompt,
             )
             result = await cohort_task.await_result()
             if isinstance(result, TableResult):
@@ -661,6 +687,8 @@ async def dedupe(
         session=session,
         input=input,
         equivalence_relation=equivalence_relation,
+        strategy=strategy,
+        strategy_prompt=strategy_prompt,
     )
     result = await cohort_task.await_result()
     if isinstance(result, TableResult):
@@ -672,6 +700,8 @@ async def dedupe_async(
     session: Session,
     input: DataFrame | UUID | TableResult,
     equivalence_relation: str,
+    strategy: Literal["identify", "select", "combine"] | None = None,
+    strategy_prompt: str | None = None,
 ) -> EveryrowTask[BaseModel]:
     """Submit a dedupe task asynchronously."""
     input_data = _prepare_table_input(input, DedupeOperationInputType1Item)
@@ -680,6 +710,8 @@ async def dedupe_async(
         input_=input_data,  # type: ignore
         equivalence_relation=equivalence_relation,
         session_id=session.session_id,
+        strategy=DedupeOperationStrategy(strategy) if strategy is not None else UNSET,
+        strategy_prompt=strategy_prompt if strategy_prompt is not None else UNSET,
     )
 
     response = await dedupe_operations_dedupe_post.asyncio(client=session.client, body=body)

tests/integration/test_dedupe.py

@@ -102,3 +102,102 @@ async def test_dedupe_unique_items_all_selected():
     assert result.data["equivalence_class_id"].nunique() == 3
     # All should be selected since they're unique
     assert result.data["selected"].all()  # pyright: ignore[reportGeneralTypeIssues]
+
+
+async def test_dedupe_identify_strategy_no_selection():
+    """Test that identify strategy clusters but does not add a 'selected' column."""
+    input_df = pd.DataFrame(
+        [
+            {"title": "Paper A - Preprint", "version": "v1"},
+            {"title": "Paper A - Published", "version": "v2"},
+            {"title": "Paper B", "version": "v1"},
+        ]
+    )
+
+    result = await dedupe(
+        equivalence_relation="""
+            Two entries are duplicates if they are versions of the same paper.
+            "Paper A - Preprint" and "Paper A - Published" are the same paper.
+        """,
+        input=input_df,
+        strategy="identify",
+    )
+
+    assert isinstance(result, TableResult)
+    assert "equivalence_class_id" in result.data.columns
+    # identify strategy should NOT add a selected column
+    assert "selected" not in result.data.columns
+
+
+async def test_dedupe_combine_strategy_creates_combined_rows():
+    """Test that combine strategy produces combined rows marked as selected."""
+    input_df = pd.DataFrame(
+        [
+            {"title": "Paper A - Preprint", "year": "2023"},
+            {"title": "Paper A - Published", "year": "2024"},
+            {"title": "Paper B", "year": "2023"},
+        ]
+    )
+
+    result = await dedupe(
+        equivalence_relation="""
+            Two entries are duplicates if they are versions of the same paper.
+            "Paper A - Preprint" and "Paper A - Published" are the same paper.
+        """,
+        input=input_df,
+        strategy="combine",
+    )
+
+    assert isinstance(result, TableResult)
+    assert "equivalence_class_id" in result.data.columns
+    assert "selected" in result.data.columns
+
+    # There should be at least one selected row
+    selected_rows = result.data[result.data["selected"] == True]  # noqa: E712
+    assert len(selected_rows) >= 1
+
+
+async def test_dedupe_select_strategy_explicit():
+    """Test that explicitly passing strategy='select' works the same as the default."""
+    input_df = pd.DataFrame(
+        [
+            {"item": "Apple"},
+            {"item": "Banana"},
+        ]
+    )
+
+    result = await dedupe(
+        equivalence_relation="Items are duplicates only if they are the exact same fruit name.",
+        input=input_df,
+        strategy="select",
+    )
+
+    assert isinstance(result, TableResult)
+    assert "equivalence_class_id" in result.data.columns
+    assert "selected" in result.data.columns
+    # All unique items should be selected
+    assert result.data["selected"].all()  # pyright: ignore[reportGeneralTypeIssues]
+
+
+async def test_dedupe_with_strategy_prompt():
+    """Test that strategy_prompt parameter is accepted."""
+    input_df = pd.DataFrame(
+        [
+            {"title": "Paper A - Preprint", "version": "v1"},
+            {"title": "Paper A - Published", "version": "v2"},
+            {"title": "Paper B", "version": "v1"},
+        ]
+    )
+
+    result = await dedupe(
+        equivalence_relation="""
+            Two entries are duplicates if they are versions of the same paper.
+            "Paper A - Preprint" and "Paper A - Published" are the same paper.
+        """,
+        input=input_df,
+        strategy="select",
+        strategy_prompt="Always prefer the published version over the preprint.",
+    )
+
+    assert isinstance(result, TableResult)
+    assert "selected" in result.data.columns