scholarcli 1.21__tar.gz → 1.23__tar.gz

{scholarcli-1.21 → scholarcli-1.23}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: scholarcli
-Version: 1.21
+Version: 1.23
 Summary: A tool for structured literature searches across bibliographic databases
 Author-email: Daniel Bosk <dbosk@kth.se>, Ric Glassey <glassey@kth.se>
 License-Expression: MIT
@@ -303,7 +303,7 @@ scholar llm classify "session-name" --count 10
 
 ### How It Works
 
-1. **Tag some papers manually** - The LLM needs examples to learn from. Review at least 5 papers with tags (themes for kept, motivations for discarded).
+1. **Tag some papers manually** (recommended) - Examples help the LLM learn your criteria. Tagging ~5 papers (themes for kept, motivations for discarded) improves quality, but classification will still run — with a warning — if you have fewer or none.
 
 2. **Set research context** (optional) - Describe your review's focus to help the LLM understand relevance criteria.
 

{scholarcli-1.21 → scholarcli-1.23}/README.md

@@ -263,7 +263,7 @@ scholar llm classify "session-name" --count 10
 
 ### How It Works
 
-1. **Tag some papers manually** - The LLM needs examples to learn from. Review at least 5 papers with tags (themes for kept, motivations for discarded).
+1. **Tag some papers manually** (recommended) - Examples help the LLM learn your criteria. Tagging ~5 papers (themes for kept, motivations for discarded) improves quality, but classification will still run — with a warning — if you have fewer or none.
 
 2. **Set research context** (optional) - Describe your review's focus to help the LLM understand relevance criteria.
 

{scholarcli-1.21 → scholarcli-1.23}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "scholarcli"
-version = "1.21"
+version = "1.23"
 description = "A tool for structured literature searches across bibliographic databases"
 authors = [{ name = "Daniel Bosk", email = "dbosk@kth.se" },
            { name = "Ric Glassey", email = "glassey@kth.se" }]

{scholarcli-1.21 → scholarcli-1.23}/src/scholar/cli.py

@@ -3450,7 +3450,8 @@ def llm_classify(
         bool,
         typer.Option(
             "--no-examples",
-            help="Run without requiring tagged examples (zero-shot).",
+            help="Skip examples entirely for zero-shot classification, "
+            "even if some exist.",
         ),
     ] = False,
     full_text: Annotated[
@@ -3465,10 +3466,12 @@ def llm_classify(
     """
     Classify pending papers using LLM.
 
-    Uses human-reviewed papers as training examples. Requires at least
-    5 tagged examples (minimum 1 kept, 1 discarded) unless --no-examples
-    is given, which runs zero-shot classification using only the research
-    context.
+    Uses human-reviewed papers as training examples when available. Tagged
+    examples are recommended (ideally 5+, with at least 1 kept and 1
+    discarded) but no longer required: if too few exist, a warning is shown
+    and classification proceeds with whatever examples are present. Pass
+    --no-examples to skip examples entirely and run zero-shot classification
+    using only the research context, even when examples are available.
 
     With --full-text, supporting passages from each paper's PDF (matched
     against the research context) are added to the prompt, so classification

{scholarcli-1.21 → scholarcli-1.23}/src/scholar/llm_review.py

@@ -537,7 +537,8 @@ def classify_papers_with_llm(
 
     This is the main entry point for LLM-assisted classification. It:
     1. Optionally gathers training examples from human-reviewed papers
-    2. Optionally validates minimum example requirements
+    2. Warns (but does not fail) if examples fall below the recommended
+       minimum, then proceeds with whatever examples exist
     3. Optionally enriches papers lacking abstracts
     4. Optionally extracts full-text supporting passages
     5. Constructs a prompt with examples and papers to classify
@@ -550,9 +551,11 @@ def classify_papers_with_llm(
         model_id: Deprecated compatibility alias for selecting the analytic model
         enrich_missing: Whether to auto-enrich papers without abstracts
         dry_run: If True, return the prompt without calling LLM
-        require_examples: If True (default), require tagged example papers
-            before classification. Set False for zero-shot classification
-            using only the research context.
+        require_examples: If True (default), gather tagged example papers
+            and use them to ground the LLM. Insufficient or missing examples
+            produce a warning, not an error. Set False to skip example
+            gathering entirely for zero-shot classification using only the
+            research context.
         use_fulltext: If True, extract supporting passages from each paper's
             full text (matched against the research context) and include them
             in the prompt, so classification can use evidence beyond the
@@ -563,18 +566,22 @@ def classify_papers_with_llm(
         LLMBatchResult with decisions, or prompt string if dry_run=True
 
     Raises:
-        ValueError: If there are no papers to classify, or if
-            require_examples=True and there are insufficient examples
+        ValueError: If there are no pending papers to classify
         ImportError: If llm package is not installed
     """
     # Gather examples (optional)
     if require_examples:
         kept_examples, discarded_examples = get_example_decisions(session)
 
-        # Validate
+        # Below-threshold examples are a quality concern, not a hard stop:
+        # warn and continue with whatever was gathered (possibly none) so
+        # classification still runs early in a review.
         is_valid, error = validate_examples(kept_examples, discarded_examples)
         if not is_valid:
-            raise ValueError(error)
+            logger.warning(
+                f"{error} Proceeding with the available examples; "
+                "classification quality may be reduced."
+            )
     else:
         kept_examples = []
         discarded_examples = []

{scholarcli-1.21 → scholarcli-1.23}/src/scholarcli.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: scholarcli
-Version: 1.21
+Version: 1.23
 Summary: A tool for structured literature searches across bibliographic databases
 Author-email: Daniel Bosk <dbosk@kth.se>, Ric Glassey <glassey@kth.se>
 License-Expression: MIT
@@ -303,7 +303,7 @@ scholar llm classify "session-name" --count 10
 
 ### How It Works
 
-1. **Tag some papers manually** - The LLM needs examples to learn from. Review at least 5 papers with tags (themes for kept, motivations for discarded).
+1. **Tag some papers manually** (recommended) - Examples help the LLM learn your criteria. Tagging ~5 papers (themes for kept, motivations for discarded) improves quality, but classification will still run — with a warning — if you have fewer or none.
 
 2. **Set research context** (optional) - Describe your review's focus to help the LLM understand relevance criteria.
 

{scholarcli-1.21 → scholarcli-1.23}/tests/test_llm_review.py

@@ -1,5 +1,6 @@
 """Tests for the LLM review module."""
 import json
+import logging
 import pytest
 from datetime import datetime
 from unittest.mock import Mock, patch, MagicMock
@@ -380,6 +381,39 @@ class TestZeroShotClassification:
 
         assert "Research Context" in prompt
         assert "Papers to Classify" in prompt
+class TestExamplesOptional:
+    """Default classification tolerates missing/insufficient examples."""
+
+    def test_classify_insufficient_examples_warns_not_raises(self, caplog):
+        """Below-threshold examples warn and still build a prompt."""
+        session = ReviewSession(
+            query="test",
+            providers=["test"],
+            timestamp=datetime.now(),
+            research_context="I am studying X.",
+        )
+        session.decisions.append(
+            ReviewDecision(
+                paper=Paper(
+                    title="Pending",
+                    authors=["A"],
+                    year=2024,
+                    abstract="Abstract.",
+                ),
+                provider="test",
+                status=DecisionStatus.PENDING,
+            )
+        )
+
+        with caplog.at_level(logging.WARNING):
+            prompt = classify_papers_with_llm(
+                session=session,
+                count=1,
+                dry_run=True,
+            )
+
+        assert "Papers to Classify" in prompt
+        assert "Proceeding with the available examples" in caplog.text
 class TestLLMInteraction:
     """Tests for LLM interaction functions."""