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.

refactor: remove redundant per-site <think>-stripping; rely on central wrapper (#4343)

LearningCircuita761417b6df6
diff --git a/changelog.d/+remove-redundant-think-stripping.bugfix.md b/changelog.d/+remove-redundant-think-stripping.bugfix.md
new file mode 100644
index 0000000000..938c421524
--- /dev/null
+++ b/changelog.d/+remove-redundant-think-stripping.bugfix.md
@@ -0,0 +1 @@
+Remove redundant per-site `<think>`-tag stripping now that `get_llm` LLMs are normalized centrally in `ProcessingLLMWrapper`. Direct `self.model.invoke(...).content` is already think-stripped, so ~25 scattered `remove_think_tags(...)` calls across the search strategies, constraint/evidence/knowledge components, and synthesis path were dropped (relying on the single central strip point), keeping explicit handling only where it is load-bearing (agent/`bind_tools` paths, injected LLMs). Empty-answer fallbacks now log a warning so a silent degrade is visible.
diff --git a/src/local_deep_research/advanced_search_system/constraints/constraint_analyzer.py b/src/local_deep_research/advanced_search_system/constraints/constraint_analyzer.py
index 507f61491f..fc8b82117e 100644
--- a/src/local_deep_research/advanced_search_system/constraints/constraint_analyzer.py
+++ b/src/local_deep_research/advanced_search_system/constraints/constraint_analyzer.py
@@ -8,7 +8,6 @@ from typing import List
from langchain_core.language_models import BaseChatModel
from loguru import logger
-from ...utilities.search_utilities import remove_think_tags
from .base_constraint import Constraint, ConstraintType
@@ -57,7 +56,7 @@ Focus on answer verification, not query parsing.
"""
response = self.model.invoke(prompt)
- content = remove_think_tags(response.content)
+ content = response.content
constraints = []
current_constraint = {}
diff --git a/src/local_deep_research/advanced_search_system/evidence/evaluator.py b/src/local_deep_research/advanced_search_system/evidence/evaluator.py
index 0c9a900a51..fca4c2d3b6 100644
--- a/src/local_deep_research/advanced_search_system/evidence/evaluator.py
+++ b/src/local_deep_research/advanced_search_system/evidence/evaluator.py
@@ -7,7 +7,6 @@ from typing import Dict
from langchain_core.language_models import BaseChatModel
from loguru import logger
-from ...utilities.search_utilities import remove_think_tags
from ..constraints.base_constraint import Constraint
from .base_evidence import Evidence, EvidenceType
@@ -59,7 +58,7 @@ QUOTE: [relevant text]
"""
response = self.model.invoke(prompt)
- content = remove_think_tags(response.content)
+ content = response.content
# Parse response
parsed = self._parse_evidence_response(content)
diff --git a/src/local_deep_research/advanced_search_system/findings/repository.py b/src/local_deep_research/advanced_search_system/findings/repository.py
index 96e4db6caf..14035f6bfe 100644
--- a/src/local_deep_research/advanced_search_system/findings/repository.py
+++ b/src/local_deep_research/advanced_search_system/findings/repository.py
@@ -8,7 +8,6 @@ from typing import Dict, List, Union
from langchain_core.documents import Document
from langchain_core.language_models import BaseLanguageModel
-from ...utilities.json_utils import get_llm_response_text
from ...utilities.search_utilities import format_findings
from .base_findings import BaseFindingsRepository
@@ -363,8 +362,12 @@ Use IEEE style citations [1], [2], etc. Never make up your own citations.
120, self.model.invoke, prompt
)
- # Normalize response (handles str/.content, strips <think> tags)
- synthesized_content = get_llm_response_text(response)
+ # <think> tags are stripped centrally in ProcessingLLMWrapper;
+ # keep only a shape-guard for non-message returns.
+ if hasattr(response, "content"):
+ synthesized_content = response.content
+ else:
+ synthesized_content = str(response)
logger.info(
f"Successfully synthesized final answer for query: '{query}'"
@@ -399,10 +402,12 @@ Use IEEE style citations [1], [2], etc. Never make up your own citations.
):
response = self.model.invoke(prompt)
- # Normalize response (handles str/.content, strips <think> tags)
- synthesized_content = get_llm_response_text(
- response
- )
+ # <think> tags are stripped centrally in ProcessingLLMWrapper;
+ # keep only a shape-guard for non-message returns.
+ if hasattr(response, "content"):
+ synthesized_content = response.content
+ else:
+ synthesized_content = str(response)
logger.info(
f"Successfully synthesized final answer for query: '{query}'"
diff --git a/src/local_deep_research/advanced_search_system/knowledge/followup_context_manager.py b/src/local_deep_research/advanced_search_system/knowledge/followup_context_manager.py
index 2e9c7d507c..31a26c0cb1 100644
--- a/src/local_deep_research/advanced_search_system/knowledge/followup_context_manager.py
+++ b/src/local_deep_research/advanced_search_system/knowledge/followup_context_manager.py
@@ -10,7 +10,6 @@ from typing import Dict, List, Any, Optional
from loguru import logger
from langchain_core.language_models.chat_models import BaseChatModel
-from ...utilities.search_utilities import remove_think_tags
class FollowUpContextHandler:
@@ -156,9 +155,7 @@ Return up to 10 most important entities, one per line.
response = self.model.invoke(prompt)
entities = [
line.strip()
- for line in remove_think_tags(response.content)
- .strip()
- .split("\n")
+ for line in response.content.strip().split("\n")
if line.strip()
]
return entities[:10]
@@ -296,7 +293,7 @@ Create a summary of {max_sentences} sentences that captures the most relevant in
try:
response = self.model.invoke(prompt)
- summary = remove_think_tags(response.content).strip()
+ summary = response.content.strip()
# Apply length constraint if specified
if max_length and len(summary) > max_length:
@@ -343,9 +340,7 @@ What specific information is missing or needs clarification? List up to 5 gaps,
response = self.model.invoke(prompt)
gaps = [
line.strip()
- for line in remove_think_tags(response.content)
- .strip()
- .split("\n")
+ for line in response.content.strip().split("\n")
if line.strip()
]
return gaps[:5]
diff --git a/src/local_deep_research/advanced_search_system/knowledge/standard_knowledge.py b/src/local_deep_research/advanced_search_system/knowledge/standard_knowledge.py
index bcb996faf8..fb9a78081f 100644
--- a/src/local_deep_research/advanced_search_system/knowledge/standard_knowledge.py
+++ b/src/local_deep_research/advanced_search_system/knowledge/standard_knowledge.py
@@ -6,7 +6,6 @@ from loguru import logger
from datetime import datetime, UTC
from typing import List
-from ...utilities.json_utils import get_llm_response_text
from .base_knowledge import BaseKnowledgeGenerator
@@ -59,7 +58,7 @@ Generate detailed knowledge that:
Format your response as a well-structured paragraph."""
- knowledge = get_llm_response_text(self.model.invoke(prompt))
+ knowledge = self.model.invoke(prompt).content
logger.info("Generated knowledge successfully")
return knowledge
@@ -90,7 +89,7 @@ Generate detailed knowledge that:
Format your response as a well-structured paragraph."""
try:
- return get_llm_response_text(self.model.invoke(prompt))
+ return self.model.invoke(prompt).content
except Exception:
logger.exception("Error generating sub-knowledge")
return ""
@@ -127,9 +126,7 @@ Accumulated Knowledge:
Compressed Knowledge:"""
try:
- compressed_knowledge = get_llm_response_text(
- self.model.invoke(prompt)
- )
+ compressed_knowledge = self.model.invoke(prompt).content
logger.info(
f"Compressed knowledge length: {len(compressed_knowledge)}"
)
diff --git a/src/local_deep_research/api/research_functions.py b/src/local_deep_research/api/research_functions.py
index 84847f08d4..7c9b335b64 100644
--- a/src/local_deep_research/api/research_functions.py
+++ b/src/local_deep_research/api/research_functions.py
@@ -16,7 +16,6 @@ from ..report_generator import IntegratedReportGenerator
from ..search_system import AdvancedSearchSystem
from ..utilities.db_utils import no_db_settings
from ..utilities.thread_context import clear_search_context, set_search_context
-from ..utilities.search_utilities import remove_think_tags
from .settings_utils import create_settings_snapshot
@@ -730,7 +729,7 @@ def analyze_documents(
logger.info(f"LLM summary generation completed in {llm_elapsed:.2f}s")
if hasattr(summary_response, "content"):
- summary = remove_think_tags(summary_response.content)
+ summary = summary_response.content
else:
summary = str(summary_response)
diff --git a/src/local_deep_research/citation_handlers/precision_extraction_handler.py b/src/local_deep_research/citation_handlers/precision_extraction_handler.py
index f173692d03..a38225882f 100644
--- a/src/local_deep_research/citation_handlers/precision_extraction_handler.py
+++ b/src/local_deep_research/citation_handlers/precision_extraction_handler.py
@@ -485,6 +485,10 @@ The answer is:"""
try:
answer = self._invoke_text(extraction_prompt)
if not answer:
+ logger.warning(
+ "Empty LLM answer in score extraction; "
+ "falling back to first matched score"
+ )
return f"{scores[0][0]}-{scores[0][1]}. {content}"
return f"{answer}. {content}"
except Exception:
@@ -511,6 +515,10 @@ The answer is:"""
try:
answer = self._invoke_text(extraction_prompt)
if not answer:
+ logger.warning(
+ "Empty LLM answer in temporal extraction; "
+ "falling back to first matched year"
+ )
return f"{years[0]}. {content}"
# Clean to just the year/date
year_match = self.answer_patterns["year"].search(answer)
@@ -542,6 +550,10 @@ The answer is:"""
try:
answer = self._invoke_text(extraction_prompt)
if not answer:
+ logger.warning(
+ "Empty LLM answer in number extraction; "
+ "falling back to first matched number"
+ )
return f"{numbers[0]}. {content}"
return f"{answer}. {content}"
except Exception:
diff --git a/src/local_deep_research/config/llm_config.py b/src/local_deep_research/config/llm_config.py
index b834b43906..953fd73692 100644
--- a/src/local_deep_research/config/llm_config.py
+++ b/src/local_deep_research/config/llm_config.py
@@ -398,6 +398,14 @@ def wrap_llm_without_think_tags(
def _normalize_response(response: Any) -> Any:
"""Strip <think> tags and normalize the response shape.
+ This is the SINGLE authorized place that strips <think> tags from a
+ fresh LLM response. Every LLM from get_llm() is wrapped here, so
+ downstream code can use ``response.content`` directly — do NOT add
+ per-site ``remove_think_tags`` / ``get_llm_response_text`` calls on
+ fresh ``invoke``/``ainvoke`` results; they are redundant and hide
+ bugs. (Exceptions: agent/bind_tools paths bypass this wrapper, and
+ injected/unwrapped LLMs — those may still need explicit handling.)
+
A message keeps its object identity (only ``.content`` is rewritten,
so ``additional_kwargs``/``reasoning_content``/``tool_calls`` survive).
A bare-string return (some providers/wrappers) is wrapped into an
diff --git a/src/local_deep_research/utilities/search_utilities.py b/src/local_deep_research/utilities/search_utilities.py
index 3adc109c19..ac94218a04 100644
--- a/src/local_deep_research/utilities/search_utilities.py
+++ b/src/local_deep_research/utilities/search_utilities.py
@@ -24,6 +24,10 @@ LANGUAGE_CODE_MAP = {
def remove_think_tags(text: str) -> str:
+ # NOTE: Fresh LLM responses from get_llm() are already <think>-stripped
+ # centrally by ProcessingLLMWrapper (config/llm_config.py). Use this only on
+ # text NOT from a fresh wrapped invoke (accumulated/concatenated text, or
+ # agent/bind_tools output that bypasses the wrapper).
# Remove paired <think>...</think> tags
text = re.sub(r"<think>.*?</think>", "", text, flags=re.DOTALL)
# Remove any orphaned opening or closing think tags
diff --git a/tests/advanced_search_system/evidence/test_evaluator.py b/tests/advanced_search_system/evidence/test_evaluator.py
index 38e72a0c36..b0fb2ebc8e 100644
--- a/tests/advanced_search_system/evidence/test_evaluator.py
+++ b/tests/advanced_search_system/evidence/test_evaluator.py
@@ -8,7 +8,7 @@ test_evaluator_integration.py:
- extract_evidence handles invalid confidence values gracefully
"""
-from unittest.mock import Mock, patch
+from unittest.mock import Mock
class TestEvidenceEvaluatorInit:
@@ -29,12 +29,7 @@ class TestEvidenceEvaluatorInit:
class TestExtractEvidence:
"""Tests for extract_evidence method."""
- @patch(
- "local_deep_research.advanced_search_system.evidence.evaluator.remove_think_tags"
- )
- def test_extract_evidence_handles_invalid_confidence(
- self, mock_remove_tags
- ):
+ def test_extract_evidence_handles_invalid_confidence(self):
"""Test that invalid confidence is handled."""
from local_deep_research.advanced_search_system.evidence.evaluator import (
EvidenceEvaluator,
@@ -51,7 +46,6 @@ TYPE: inference
SOURCE: Analysis
CONFIDENCE: not_a_number"""
)
- mock_remove_tags.side_effect = lambda x: x
evaluator = EvidenceEvaluator(mock_model)
constraint = Constraint(
diff --git a/tests/advanced_search_system/findings/test_findings_repository.py b/tests/advanced_search_system/findings/test_findings_repository.py
index 5f356b7838..1d9c1afef6 100644
--- a/tests/advanced_search_system/findings/test_findings_repository.py
+++ b/tests/advanced_search_system/findings/test_findings_repository.py
@@ -129,28 +129,6 @@ class TestSynthesizeFindings:
assert result == "Synthesized answer"
- def test_strips_think_tags_from_synthesis(self):
- """synthesize_findings strips <think> reasoning from the final answer."""
- from local_deep_research.advanced_search_system.findings.repository import (
- FindingsRepository,
- )
-
- mock_model = Mock()
- mock_model.invoke.return_value = Mock(
- content="<think>weighing the evidence</think>The final answer is X [1]."
- )
-
- repo = FindingsRepository(mock_model)
- result = repo.synthesize_findings(
- query="Test query",
- sub_queries=[],
- findings=[{"content": "Finding 1"}],
- )
-
- assert "<think>" not in result
- assert "weighing the evidence" not in result
- assert result == "The final answer is X [1]."
-
def test_synthesizes_string_response(self):
"""synthesize_findings handles a raw string model return (no .content)."""
from local_deep_research.advanced_search_system.findings.repository import (
diff --git a/tests/advanced_search_system/knowledge/test_standard_knowledge.py b/tests/advanced_search_system/knowledge/test_standard_knowledge.py
index fa7bea1052..8f82712c7f 100644
--- a/tests/advanced_search_system/knowledge/test_standard_knowledge.py
+++ b/tests/advanced_search_system/knowledge/test_standard_knowledge.py
@@ -59,43 +59,6 @@ class TestGenerateKnowledge:
assert result == "Generated knowledge about topic"
mock_model.invoke.assert_called_once()
- def test_strips_think_tags_from_knowledge(self):
- """generate_knowledge strips <think> reasoning from the output."""
- from local_deep_research.advanced_search_system.knowledge.standard_knowledge import (
- StandardKnowledge,
- )
-
- mock_model = Mock()
- mock_model.invoke.return_value = Mock(
- content="<think>planning the answer</think>Quantum computing uses qubits."
- )
-
- knowledge = StandardKnowledge(mock_model)
- result = knowledge.generate_knowledge("What is quantum computing?")
-
- assert "<think>" not in result
- assert "planning the answer" not in result
- assert result == "Quantum computing uses qubits."
-
- def test_compress_knowledge_strips_think_tags(self):
- """compress_knowledge strips <think> reasoning from the output."""
- from local_deep_research.advanced_search_system.knowledge.standard_knowledge import (
- StandardKnowledge,
- )
-
- mock_model = Mock()
- mock_model.invoke.return_value = Mock(
- content="<think>condensing</think>Compressed facts [1]."
- )
-
- knowledge = StandardKnowledge(mock_model)
- result = knowledge.compress_knowledge(
- "Long accumulated knowledge", "query", []
- )
-
- assert "<think>" not in result
- assert result == "Compressed facts [1]."
-
def test_includes_query_in_prompt(self):
"""Test that query is included in prompt."""
from local_deep_research.advanced_search_system.knowledge.standard_knowledge import (