ndpvt-web/diverge-diversity-enhanced-rag-open-ended

DIVERGE: Diversity-Enhanced RAG for Open-Ended Information Seeking

This skill enables Claude to implement the DIVERGE agentic RAG framework, which generates multiple diverse yet high-quality responses to open-ended queries instead of collapsing to a single dominant answer. Standard RAG pipelines assume one correct answer per query; DIVERGE breaks this assumption through reflection-guided viewpoint generation, diversity-aware retrieval re-ranking, and a memory buffer that tracks explored perspectives across iterations. Use this when building search systems, recommendation engines, brainstorming tools, or any pipeline where query answers are legitimately plural.

When to Use

• When the user asks to build a RAG pipeline that handles open-ended questions with multiple valid answers (e.g., "What are good date night restaurants?" or "How can I improve my sleep?")
• When implementing a search or Q&A system that should surface diverse perspectives rather than N near-duplicate results
• When the user wants to brainstorm or ideate by retrieving and synthesizing multiple distinct viewpoints from a corpus
• When building a recommendation system where variety matters (travel suggestions, product alternatives, career paths)
• When the user notices their RAG system keeps returning the same angle regardless of how many documents are retrieved
• When implementing fair and inclusive information access where minority viewpoints should not be suppressed by majority consensus

Key Technique

The core insight: Standard RAG systems underutilize retrieved context diversity. Even when retrieval returns documents covering multiple perspectives, the LLM generation step collapses them into a single dominant response. Simply increasing retrieval diversity does not fix this — the bottleneck is in generation, not retrieval.

DIVERGE solves this with three interlocking mechanisms. First, reflection-guided viewpoint generation: after producing an initial response, the system extracts the underlying viewpoints (e.g., "budget-focused", "health-focused", "convenience-focused"), then explicitly reflects on which perspectives remain unexplored and generates a new viewpoint to target. Second, viewpoint-aware diversity retrieval: each iteration retrieves documents using a modified Maximal Marginal Relevance (MMR) score that penalizes similarity to previously retrieved contexts across all prior iterations, not just the current one. The scoring formula balances relevance to the viewpoint-conditioned query against cross-iteration and within-iteration diversity. Third, memory-augmented iterative refinement: a lightweight buffer stores (query, documents, viewpoint, answer) tuples from each iteration, giving the system full awareness of what has already been explored.

The diversity-quality trade-off is measured using semantic diversity (pairwise cosine distance between response embeddings), viewpoint diversity (unique atomic claims across responses), and quality (factual accuracy, evidence support, consistency, relevance). The unified score UQ^D is the harmonic mean of normalized quality and diversity, providing a single metric for the trade-off.

Step-by-Step Workflow

1. Classify the query as open-ended or closed. Check whether the query has a single factual answer ("What year was Python created?") or multiple valid answers ("What's the best way to learn Python?"). Only apply DIVERGE to open-ended queries — closed queries should use standard RAG.

2. Perform initial retrieval and generation. Run a standard retrieval step against your corpus (vector search, BM25, or hybrid). Generate the first response using the retrieved documents. This becomes iteration t=1.

3. Extract viewpoints from the initial response. Prompt the LLM to decompose the response into its underlying viewpoints — the distinct perspectives, assumptions, or angles it covers. Output these as a structured list (e.g., ["cost-effectiveness", "ease of use", "community support"]).

4. Initialize the memory buffer. Store the first iteration as a tuple: {query, retrieved_docs, viewpoints, answer}. This buffer persists across all iterations and is the system's "explored territory" map.

5. Reflect to identify an unexplored viewpoint. Prompt the LLM with the memory buffer contents and ask: "Given the query and the viewpoints already explored, identify a new meaningful perspective that has not been covered. Avoid restating existing viewpoints at a higher abstraction level." The reflection must produce a concrete, novel viewpoint.

6. Construct a viewpoint-conditioned query. Rewrite the original query to emphasize the new viewpoint. For example, if the original query is "How can I improve my sleep?" and the new viewpoint is "environmental factors", the conditioned query becomes "How can I improve my sleep by changing my environment?"

7. Retrieve with diversity-aware re-ranking. Retrieve candidate documents using the conditioned query, then re-rank using the modified MMR formula: score(d) = alpha * Rel(d, q_t) - beta * max_similarity_to_memory - (1-alpha) * max_similarity_within_batch. This ensures new documents are relevant to the target viewpoint while being distinct from all previously retrieved contexts.

8. Generate a viewpoint-conditioned response. Prompt the LLM to answer the original query from the specific viewpoint, grounded in the newly retrieved documents. Include a refinement instruction: "Ensure the response addresses the original question while focusing on the [viewpoint] perspective."

9. Update the memory buffer and repeat. Append the new iteration's tuple to memory. Return to step 5 and repeat for K iterations total (typically K=5-10 depending on the desired diversity breadth and compute budget).

10. Aggregate and deduplicate the final response set. Collect all K responses. Optionally compute pairwise semantic similarity and merge near-duplicates. Present the responses organized by viewpoint, or synthesize them into a single comprehensive answer that explicitly labels each perspective.

Concrete Examples

Example 1: Building a diverse travel recommendation RAG

User: "I'm building a travel Q&A chatbot. When someone asks 'Where should I travel in Southeast Asia?', my RAG keeps returning the same Bangkok/Bali recommendations. Help me implement DIVERGE to get diverse suggestions."

Approach:

1. Classify as open-ended (multiple valid destinations, no single correct answer)
2. Initial retrieval + generation produces: "Bangkok for street food, Bali for beaches"
3. Extract viewpoints: ["popular tourist destinations", "food-focused travel", "beach vacations"]
4. Store in memory buffer
5. Reflect: "Unexplored viewpoints: budget backpacking, cultural immersion, adventure/nature, off-the-beaten-path"
6. Conditioned query for iteration 2: "Where should I travel in Southeast Asia for cultural immersion experiences?"
7. Diversity-aware retrieval pulls documents about Luang Prabang, Yogyakarta, Hoi An — penalizing Bangkok/Bali docs
8. Generate: "For deep cultural immersion, consider Luang Prabang in Laos..."
9. Repeat for "adventure/nature" (retrieves Borneo, Komodo, northern Vietnam), "budget backpacking" (retrieves Cambodia, Myanmar), etc.

Output (5 iterations):

Viewpoint 1 — Popular highlights: Bangkok, Bali, Singapore
Viewpoint 2 — Cultural immersion: Luang Prabang, Yogyakarta, Hoi An
Viewpoint 3 — Adventure & nature: Borneo rainforest, Komodo, Ha Giang Loop
Viewpoint 4 — Budget backpacking: Cambodia, Myanmar, rural Philippines
Viewpoint 5 — Wellness & retreat: Ubud, Chiang Mai, Koh Phangan

Example 2: Diverse code architecture suggestions

User: "I want to use DIVERGE to give developers multiple architectural options when they ask 'How should I structure my backend API?'"

Approach:

1. Open-ended: multiple valid architectures exist
2. Initial generation: "Use a REST API with Express.js and MVC pattern"
3. Extract viewpoints: ["REST paradigm", "Node.js ecosystem", "MVC pattern"]
4. Reflect: unexplored — GraphQL, microservices vs monolith, serverless, event-driven
5. Iteration 2 conditioned on "event-driven architecture": retrieves docs on CQRS, message queues
6. Iteration 3 conditioned on "serverless": retrieves docs on AWS Lambda, API Gateway patterns
7. Iteration 4 conditioned on "type-safe API design": retrieves docs on tRPC, gRPC

Output:

Option A — Classic REST + MVC: Express.js with controllers/services/models separation
Option B — Event-driven CQRS: Commands and queries separated, message bus for async ops
Option C — Serverless functions: AWS Lambda per endpoint, API Gateway routing
Option D — Type-safe RPC: tRPC or gRPC with generated clients, no REST overhead
Each option includes: when to use it, trade-offs, and example folder structure.

Example 3: Implementing the diversity-aware retrieval re-ranker

User: "Show me how to implement the DIVERGE MMR re-ranking step in Python."

import numpy as np
from typing import List, Dict

def diverge_mmr_rerank(
    candidates: List[Dict],        # [{id, embedding, text}, ...]
    query_embedding: np.ndarray,    # embedding of viewpoint-conditioned query
    memory_embeddings: List[np.ndarray],  # embeddings from all prior iterations
    alpha: float = 0.6,            # relevance vs within-batch diversity
    beta: float = 0.3,             # cross-iteration diversity penalty
    top_k: int = 5
) -> List[Dict]:
    """Re-rank candidates using DIVERGE's diversity-aware MMR."""
    selected = []
    remaining = list(range(len(candidates)))

    for _ in range(min(top_k, len(candidates))):
        best_score, best_idx = -float('inf'), -1
        for i in remaining:
            emb = candidates[i]['embedding']

            # Relevance to viewpoint-conditioned query
            relevance = np.dot(emb, query_embedding) / (
                np.linalg.norm(emb) * np.linalg.norm(query_embedding)
            )

            # Cross-iteration diversity: penalize similarity to memory
            cross_div = 0.0
            if memory_embeddings:
                sims = [np.dot(emb, m) / (np.linalg.norm(emb) * np.linalg.norm(m))
                        for m in memory_embeddings]
                cross_div = max(sims)

            # Within-iteration diversity: penalize similarity to already selected
            within_div = 0.0
            if selected:
                sel_embs = [candidates[s]['embedding'] for s in selected]
                within_sims = [np.dot(emb, s) / (np.linalg.norm(emb) * np.linalg.norm(s))
                               for s in sel_embs]
                within_div = max(within_sims)

            score = alpha * relevance - beta * cross_div - (1 - alpha) * within_div
            if score > best_score:
                best_score, best_idx = score, i

        selected.append(best_idx)
        remaining.remove(best_idx)

    return [candidates[i] for i in selected]

Best Practices

• Do: Extract viewpoints as concrete, distinct perspectives — not vague abstractions. "Budget-conscious approach" is good; "another way to think about it" is not.
• Do: Tune the alpha/beta parameters in the MMR re-ranker based on your domain. Higher beta (0.3-0.5) for creative tasks where diversity matters most; lower beta (0.1-0.2) for factual domains where relevance must dominate.
• Do: Include a refinement check in each iteration that validates the response still answers the original query, not just the conditioned sub-query.
• Do: Set K (iteration count) based on the breadth of the topic. Simple questions may need K=3-4; broad topics like "how to start a business" may warrant K=8-10.
• Avoid: Applying DIVERGE to factual, single-answer queries. "What is the capital of France?" does not benefit from diverse viewpoints — it adds hallucination risk.
• Avoid: Letting viewpoint reflection produce increasingly abstract or meta perspectives. If iteration 5 produces "a philosophical lens" when the query is about restaurant recommendations, the reflection prompt needs stronger grounding constraints.

Error Handling

• Viewpoint saturation: If reflection cannot produce a genuinely new viewpoint, stop early rather than force repetitive iterations. Detect this by checking semantic similarity between the new viewpoint and all prior ones — if cosine similarity > 0.85, terminate.
• Quality degradation in later iterations: Monitor response quality scores across iterations. If quality drops below a threshold (e.g., the response becomes speculative or unsupported), discard that iteration and retry with a different viewpoint.
• Retrieval returns no diverse documents: If the corpus is narrow, the diversity-aware retrieval may struggle. Fall back to standard MMR or increase the candidate pool size before re-ranking.
• Memory buffer grows too large: For very high K values, summarize older memory entries rather than passing raw tuples. Keep the most recent 3-4 iterations in full detail and compress earlier ones to viewpoint labels only.

Limitations

• DIVERGE adds latency proportional to K iterations — each iteration requires a retrieval step and an LLM generation. For real-time applications, consider running iterations in parallel where the corpus allows independent retrieval.
• The framework assumes the underlying corpus actually contains diverse perspectives. If your document collection is homogeneous (e.g., all from one source), DIVERGE cannot manufacture diversity that isn't there.
• Viewpoint extraction quality depends on the LLM's ability to identify implicit perspectives. For highly technical or niche domains, you may need domain-specific viewpoint taxonomies rather than open-ended extraction.
• The approach is designed for open-ended information seeking. Applying it to structured data queries, code generation with a single correct output, or mathematical problems will not yield meaningful benefits.

Reference

DIVERGE: Diversity-Enhanced RAG for Open-Ended Information Seeking — Hu, Tandon, Arora (2026). Focus on Section 3 (the DIVERGE framework architecture), Section 4.2 (diversity-aware MMR formula), and Section 5 (the UQ^D unified diversity-quality metric).