Began deprecation cycle of exclude_text_filter (#994)

Author: jamesbraza

paperqa/docs.py

@@ -605,28 +605,32 @@ async def aget_evidence(
         if summary_llm_model is None:
             summary_llm_model = evidence_settings.get_summary_llm()
 
-        exclude_text_filter = exclude_text_filter or set()
-        exclude_text_filter |= {c.text.name for c in session.contexts}
-
-        _k = answer_config.evidence_k
-        if exclude_text_filter:
-            # Increase k to retrieve so we have enough to down-select after retrieval
-            _k += len(exclude_text_filter)
+        if exclude_text_filter is not None:
+            text_name = Text.__name__
+            warnings.warn(
+                (
+                    "The 'exclude_text_filter' argument did not work as intended"
+                    f" due to a mix-up in excluding {text_name}.name vs {text_name}."
+                    f" This bug enabled us to have 2+ contexts per {text_name}, so to"
+                    " first-class that capability and simplify our implementation,"
+                    " we're removing the 'exclude_text_filter' argument."
+                    " This deprecation will conclude in version 6"
+                ),
+                category=DeprecationWarning,
+                stacklevel=2,
+            )
 
         if answer_config.evidence_retrieval:
             matches = await self.retrieve_texts(
                 session.question,
-                _k,
+                answer_config.evidence_k,
                 evidence_settings,
                 embedding_model,
                 partitioning_fn=partitioning_fn,
             )
         else:
             matches = self.texts
 
-        if exclude_text_filter:
-            matches = [m for m in matches if m.text not in exclude_text_filter]
-
         matches = (
             matches[: answer_config.evidence_k]
             if answer_config.evidence_retrieval

tests/test_paperqa.py

@@ -458,7 +458,7 @@ async def test_docs_lifecycle(subtests: SubTests, stub_data_dir: Path) -> None:
 
 
 @pytest.mark.asyncio
-async def test_evidence(docs_fixture) -> None:
+async def test_evidence(docs_fixture: Docs) -> None:
     debug_settings = Settings.from_name("debug")
     evidence = (
         await docs_fixture.aget_evidence(
@@ -470,6 +470,19 @@ async def test_evidence(docs_fixture) -> None:
     assert len({e.context for e in evidence}) == len(
         evidence
     ), "Expected unique contexts"
+    texts = {c.text for c in evidence}
+    assert texts, "Below assertions require at least one text to be used"
+
+    # Okay, let's check we can get other evidence using the same underlying sources
+    other_evidence = (
+        await docs_fixture.aget_evidence(
+            PQASession(question="What is an acronym for explainable AI?"),
+            settings=debug_settings,
+        )
+    ).contexts
+    assert texts.intersection(
+        {c.text for c in other_evidence}
+    ), "We should be able to reuse sources across evidence calls"
 
 
 @pytest.mark.asyncio