deep-search-research-agent

Problem

Research agents often provide incomplete answers to complex questions because they stop searching too early or fail to identify gaps in their knowledge. A single query rarely captures all relevant information, and agents need strategies to recognize when they have insufficient information and conduct targeted follow-up searches.

Solution

Implement a DeepSearchQA framework where agents:

1. Maintain Search State: Track what has been searched, what answers have been found, and what gaps remain
2. Identify Gaps: Use gap detection to recognize incomplete coverage of question aspects
3. Iterate on Queries: Generate targeted follow-up searches based on identified gaps
4. Validate Comprehensiveness: Check that answers cover multiple perspectives and dimensions of the question
5. Stop Strategically: Determine when additional searches provide diminishing returns

When to Use

• Building research assistants for complex multi-faceted questions
• Question-answering systems requiring comprehensive coverage
• Literature review and synthesis tasks
• Competitive analysis and market research automation
• Policy analysis and decision-support systems

When NOT to Use

• Simple factual lookup questions (single search suffices)
• Time-sensitive applications (iterative searching adds latency)
• Narrow technical queries with definitive answers
• Constrained search budgets or API rate limits

Implementation

Step 1: Build Search State Management

Track what has been discovered and what gaps remain.

class SearchState:
    """Maintain comprehensive search state"""

    def __init__(self, query):
        self.original_query = query
        self.search_history = []
        self.discovered_answers = []
        self.known_gaps = set()
        self.relevant_subtopics = set()
        self.coverage_map = {}

    def record_search(self, query, results):
        """Log search execution and results"""
        search_record = {
            "query": query,
            "num_results": len(results),
            "timestamp": datetime.now(),
            "results_summary": self.summarize_results(results)
        }
        self.search_history.append(search_record)

        # Extract answers and update coverage
        for result in results:
            answer = self.extract_answer(result)
            if answer not in self.discovered_answers:
                self.discovered_answers.append(answer)

    def update_coverage_map(self):
        """Map which question dimensions are covered"""
        self.coverage_map = {
            "definition": self.covers_aspect("what_is", self.discovered_answers),
            "how": self.covers_aspect("how_does", self.discovered_answers),
            "why": self.covers_aspect("why", self.discovered_answers),
            "examples": self.covers_aspect("examples", self.discovered_answers),
            "exceptions": self.covers_aspect("limitations", self.discovered_answers),
            "recent_developments": self.covers_aspect("recent", self.discovered_answers)
        }

    def identify_gaps(self):
        """Find aspects of question with insufficient coverage"""
        self.update_coverage_map()
        gaps = [aspect for aspect, covered in self.coverage_map.items() if not covered]
        self.known_gaps = set(gaps)
        return gaps

    def get_uncovered_subtopics(self):
        """Extract entities and concepts not yet fully explored"""
        all_entities = self.extract_all_entities(self.discovered_answers)
        explored = {e for e in all_entities if self.has_sufficient_coverage(e)}
        return all_entities - explored

Step 2: Implement Gap-Driven Query Generation

Create follow-up queries based on identified gaps.

def generate_follow_up_queries(state):
    """
    Generate targeted follow-up searches to address gaps
    """
    gaps = state.identify_gaps()
    uncovered_subtopics = state.get_uncovered_subtopics()

    follow_ups = []

    # For each gap, create targeted query
    for gap in gaps:
        if gap == "definition":
            query = f"What is {state.original_query}? Definition and meaning"
        elif gap == "how":
            query = f"How does {state.original_query} work? Mechanism and process"
        elif gap == "why":
            query = f"Why {state.original_query}? Reasons and motivations"
        elif gap == "examples":
            query = f"{state.original_query} examples use cases real world applications"
        elif gap == "exceptions":
            query = f"{state.original_query} limitations edge cases exceptions"
        elif gap == "recent_developments":
            query = f"{state.original_query} recent advances 2024 2025"

        follow_ups.append({
            "query": query,
            "gap_addressed": gap,
            "priority": self.compute_gap_priority(gap, state)
        })

    # Explore uncovered subtopics
    for subtopic in list(uncovered_subtopics)[:3]:  # Top 3 subtopics
        query = f"{state.original_query} {subtopic} detailed explanation"
        follow_ups.append({
            "query": query,
            "gap_addressed": f"subtopic_{subtopic}",
            "priority": 0.5
        })

    # Sort by priority (descending)
    follow_ups.sort(key=lambda x: x["priority"], reverse=True)
    return follow_ups

Step 3: Build Comprehensiveness Validator

Evaluate when answer coverage is sufficient.

class ComprehensivenessValidator:
    """Determine if search has yielded comprehensive answers"""

    def __init__(self, target_coverage=0.85):
        self.target_coverage = target_coverage

    def compute_coverage_score(self, state):
        """
        Score how comprehensively the answer covers the question.
        Considers: breadth of dimensions, depth per dimension, diversity of sources
        """
        state.update_coverage_map()

        # Dimension coverage: what % of question aspects are addressed
        dimensions = state.coverage_map.values()
        dimension_score = sum(dimensions) / len(dimensions)

        # Depth score: how many answers/perspectives per dimension
        answer_count = len(state.discovered_answers)
        depth_score = min(answer_count / 5, 1.0)  # Normalize to 5+ answers

        # Diversity score: sources/perspectives vary
        source_diversity = self.measure_source_diversity(state.discovered_answers)
        diversity_score = source_diversity

        # Composite score: weighted combination
        coverage_score = (
            0.4 * dimension_score +
            0.35 * depth_score +
            0.25 * diversity_score
        )

        return coverage_score

    def should_continue_searching(self, state, search_count):
        """
        Determine if more searches are needed
        """
        coverage = self.compute_coverage_score(state)
        gaps = state.identify_gaps()

        # Stop if coverage is high enough
        if coverage >= self.target_coverage:
            return False

        # Stop if too many searches already
        if search_count > 10:
            return False

        # Stop if diminishing returns (last 2 searches added <5% new info)
        if len(state.search_history) >= 3:
            recent_gain = self.compute_recent_information_gain(state, window=2)
            if recent_gain < 0.05:
                return False

        return True

    def compute_recent_information_gain(self, state, window=2):
        """Measure how much new information recent searches added"""
        if len(state.search_history) < window:
            return 1.0

        recent_answers = set()
        for record in state.search_history[-window:]:
            recent_answers.update(record["results_summary"])

        total_answers = set(state.discovered_answers)
        gain = len(recent_answers) / max(len(total_answers), 1)
        return gain

Step 4: Orchestrate the Deep Search Loop

Implement the main agent search iteration.

def deep_search_agent(question, max_searches=10):
    """
    Main agent loop: search -> evaluate -> identify gaps -> search again
    """
    state = SearchState(question)
    validator = ComprehensivenessValidator(target_coverage=0.8)

    search_count = 0

    while search_count < max_searches:
        if search_count == 0:
            # Initial search with original question
            query = question
        else:
            # Generate gap-driven follow-up queries
            follow_ups = generate_follow_up_queries(state)
            if not follow_ups:
                break

            query = follow_ups[0]["query"]

        # Execute search
        results = execute_search(query)
        state.record_search(query, results)
        search_count += 1

        # Check if comprehensive coverage achieved
        coverage = validator.compute_coverage_score(state)

        print(f"Search {search_count}: Coverage={coverage:.2%}, Gaps={len(state.known_gaps)}")

        # Decide whether to continue
        if not validator.should_continue_searching(state, search_count):
            break

    # Synthesize final answer
    final_answer = synthesize_answer(state.discovered_answers, state.coverage_map)

    return {
        "answer": final_answer,
        "coverage_score": validator.compute_coverage_score(state),
        "searches_conducted": search_count,
        "answer_count": len(state.discovered_answers),
        "dimensions_covered": sum(state.coverage_map.values())
    }

Step 5: Answer Synthesis and Validation

Combine discovered answers into coherent comprehensive response.

def synthesize_answer(discovered_answers, coverage_map):
    """
    Organize discovered answers by dimension (what, how, why, examples, etc.)
    """
    synthesis = {
        "summary": "",
        "dimensions": {}
    }

    # Organize by dimension
    for dimension, covered in coverage_map.items():
        if covered:
            relevant_answers = [
                a for a in discovered_answers
                if categorizes_as_dimension(a, dimension)
            ]
            synthesis["dimensions"][dimension] = {
                "answers": relevant_answers,
                "synthesized": synthesize_dimension(relevant_answers)
            }

    # Create overview
    synthesis["summary"] = create_overview(synthesis["dimensions"])

    return synthesis

Key Strategies

• Multi-Dimensional Coverage: Ensure answers cover what, how, why, examples, limitations, recent work
• Iterative Refinement: Use gap detection to guide follow-up searches
• Diminishing Returns Detection: Stop searching when new information becomes sparse
• Diversity Over Quantity: Prioritize varied sources and perspectives over raw answer count

Success Indicators

• Coverage score >= 0.8 across multiple dimensions
• Answer includes definitions, mechanisms, applications, and limitations
• Multiple independent sources support key claims
• Recent developments and current state of topic included

References

• arXiv:2601.20975: DeepSearchQA framework for comprehensive research
• Designed for question-answering agents requiring thorough coverage