Skip to content

~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.

fix(strategy): do not search a long query verbatim in focused iteration (#6384)

Lukasae13ec1685df
diff --git a/src/local_deep_research/advanced_search_system/strategies/focused_iteration_strategy.py b/src/local_deep_research/advanced_search_system/strategies/focused_iteration_strategy.py
index e38f621e62..3557432631 100644
--- a/src/local_deep_research/advanced_search_system/strategies/focused_iteration_strategy.py
+++ b/src/local_deep_research/advanced_search_system/strategies/focused_iteration_strategy.py
@@ -220,11 +220,25 @@ class FocusedIterationStrategy(BaseSearchStrategy):
questions_by_iteration=self.questions_by_iteration,
)
- # Always include original query in first iteration, but respect question limit
+ # Include the original query verbatim in the first iteration so an
+ # ambiguous topic still gets one search that hits it. A long query
+ # (a detailed-report subsection prompt embeds the whole task) is a
+ # research brief, not a search string: sent literally it full-text
+ # matches stray words such as an author surname (#6383). Same guard
+ # and setting as the source-based strategy.
if iteration == 1 and query not in questions:
- questions = [query] + questions
- # Trim to respect questions_per_iteration limit
- questions = questions[: self.questions_per_iteration]
+ max_query_length = self.get_setting(
+ "app.max_user_query_length", 300
+ )
+ if len(query.strip()) > max_query_length:
+ logger.warning(
+ f"Long user query detected ({len(query.strip())} chars > {max_query_length} limit), "
+ "using LLM questions only for search"
+ )
+ else:
+ questions = [query] + questions
+ # Trim to respect questions_per_iteration limit
+ questions = questions[: self.questions_per_iteration]
self.questions_by_iteration[iteration] = questions
logger.info(f"Iteration {iteration} questions: {questions}")
diff --git a/tests/advanced_search_system/strategies/test_focused_iteration_strategy.py b/tests/advanced_search_system/strategies/test_focused_iteration_strategy.py
index 42f041aa3b..bf996886d1 100644
--- a/tests/advanced_search_system/strategies/test_focused_iteration_strategy.py
+++ b/tests/advanced_search_system/strategies/test_focused_iteration_strategy.py
@@ -741,6 +741,78 @@ class TestAnalyzeTopic:
first_iter_questions = strategy.questions_by_iteration[1]
assert first_iter_questions[0] == "my original query"
+ def test_long_original_query_is_not_searched_verbatim(self):
+ """A query longer than app.max_user_query_length (a detailed-report
+ subsection prompt embeds the whole task) is not sent to the search
+ engine as a literal string; only the LLM questions are searched (#6383)."""
+ from local_deep_research.advanced_search_system.strategies.focused_iteration_strategy import (
+ FocusedIterationStrategy,
+ )
+
+ mock_search = Mock()
+ mock_search.run.return_value = []
+ long_query = (
+ "Research task: Create content for subsection 'History' in a report "
+ "about 'Hay et al. (2018) and juvenile justice'. " * 20
+ )
+ assert len(long_query) > 300
+
+ strategy = FocusedIterationStrategy(
+ model=Mock(),
+ search=mock_search,
+ max_iterations=1,
+ questions_per_iteration=3,
+ use_browsecomp_optimization=False,
+ )
+
+ with patch.object(
+ strategy.question_generator,
+ "generate_questions",
+ return_value=["Other Q1", "Other Q2"],
+ ):
+ with patch.object(
+ strategy.citation_handler,
+ "analyze_followup",
+ return_value={"content": "Done", "documents": []},
+ ):
+ strategy.analyze_topic(long_query)
+
+ assert strategy.questions_by_iteration[1] == ["Other Q1", "Other Q2"]
+ searched = [call.args[0] for call in mock_search.run.call_args_list]
+ assert long_query not in searched
+ assert searched == ["Other Q1", "Other Q2"]
+
+ def test_query_length_limit_follows_the_setting(self):
+ """The cutoff is app.max_user_query_length, shared with the source-based strategy."""
+ from local_deep_research.advanced_search_system.strategies.focused_iteration_strategy import (
+ FocusedIterationStrategy,
+ )
+
+ mock_search = Mock()
+ mock_search.run.return_value = []
+ strategy = FocusedIterationStrategy(
+ model=Mock(),
+ search=mock_search,
+ max_iterations=1,
+ questions_per_iteration=3,
+ use_browsecomp_optimization=False,
+ settings_snapshot={"app.max_user_query_length": 10},
+ )
+
+ with patch.object(
+ strategy.question_generator,
+ "generate_questions",
+ return_value=["Other Q1"],
+ ):
+ with patch.object(
+ strategy.citation_handler,
+ "analyze_followup",
+ return_value={"content": "Done", "documents": []},
+ ):
+ strategy.analyze_topic("short query") # 11 chars > 10
+
+ assert strategy.questions_by_iteration[1] == ["Other Q1"]
+
def test_exception_returns_error_response(self):
"""Exception during analysis returns error response."""
from local_deep_research.advanced_search_system.strategies.focused_iteration_strategy import (