The Problem with Vanilla RAG
You have built a RAG system. It works great for simple questions, but then someone asks: How does Anthropic's approach to AI safety differ from OpenAI's? What are the implications for the industry?
In such a case, your system retrieves a few chunks, generates a response, and...it's shallow. It missed half the question. It didn't connect the dots.
This is the fundamental limitation of single-shot retrieval. Complex questions require reasoning—breaking problems down, retrieving iteratively, and synthesizing across multiple sources. They require your RAG system to think.
Enter Chain of Thought (CoT) for RAG.
What is Chain of Thought Retrieval?
Chain of Thought prompting, introduced by Google researchers in 2022, showed that language models perform dramatically better on complex tasks when they "show their work"—reasoning step by step rather than jumping to answers.
The insight for RAG systems: don't just retrieve once and generate. Reason about what you need, retrieve it, reason about what's still missing, retrieve again, and synthesize.
Instead of: Query → Retrieve → Generate
We get: Query → Think → Retrieve → Think → Retrieve → ... → Synthesize
This simple shift unlocks multi-hop reasoning, self-correction, and dramatically better answers on complex questions.
Seven Approaches to CoT-RAG
The approaches listed below are mapped from the simplest to the most complex.
1. Query Decomposition: Plan First, Execute in Parallel
The simplest approach: break the question into sub-questions upfront, retrieve for each (in parallel), then synthesize.
How it works:
def decomposition_rag(query: str):
# Step 1: Break it down
sub_queries = llm.generate(f"""
Break this into 2-4 independent sub-questions:
{query}
""")
# Step 2: Retrieve in parallel
results = parallel_retrieve(sub_queries)
# Step 3: Synthesize
return llm.generate(f"""
Original question: {query}
Research: {results}
Synthesize a complete answer.
""")
Example in action:
User: "Compare Tesla and Ford's EV strategies"
Decomposed:
→ "Tesla electric vehicle strategy 2024"
→ "Ford electric vehicle strategy 2024"
→ "EV market competitive landscape"
[Parallel retrieval for all three]
[Synthesize into comparative analysis]
When to use it: Predictable queries where you can anticipate the sub-questions. Great for comparison questions, multi-part requests, and research tasks.
Trade-offs: Fast (parallel retrieval) but inflexible. If your decomposition is wrong, you can't adapt mid-flight.
2. ReAct: Reasoning and Acting in a Loop
ReAct (Reasoning + Acting) interleaves thinking with action. The model reasons about what to do, takes an action (like searching), observes the result, and repeats.
The pattern:
Thought → Action → Observation → Thought → Action → Observation → ... → Answer
Implementation:
def react_rag(query: str, max_steps: int = 5):
messages = [{"role": "user", "content": f"""
Answer this question: {query}
Think step by step. Format each step as:
Thought: <reasoning about what to do>
Action: search[<query>] OR answer[<final answer>]
"""}]
for _ in range(max_steps):
response = llm.generate(messages)
thought, action = parse_response(response)
if action.startswith("answer"):
return extract_answer(action)
# Execute search and add observation
results = retriever.search(extract_query(action))
messages.append({"role": "assistant", "content": response})
messages.append({"role": "user", "content": f"Observation: {results}"})
return "Could not find sufficient information"
Example trace:
Thought: This is a comparative question. I need information about
Anthropic's safety approach first.
Action: search[Anthropic AI safety techniques]
Observation: [Chunks about Constitutional AI, RLHF, interpretability...]
Thought: Good, I learned about Constitutional AI. Now I need
OpenAI's approach for comparison.
Action: search[OpenAI safety alignment methods]
Observation: [Chunks about RLHF, red teaming, iterative deployment...]
Thought: I have both perspectives. Key differences are emerging:
Anthropic focuses on Constitutional AI with explicit principles,
while OpenAI emphasizes iterative deployment. I can now synthesize.
Action: answer[Anthropic and OpenAI share RLHF but differ in key ways...]
When to use it: Complex, multi-hop questions where you can't predict what information you'll need. Great when adaptability matters more than speed.
Trade-offs: Highly adaptive and interpretable, but higher latency due to sequential LLM calls. Can also "over-search" if not carefully constrained.
Similar to "Reasoning and Acting" (ReAct), but the model explicitly asks and answers intermediate questions. The structure is more rigid but often easier to implement.
The pattern:
Question: [complex query]
Are follow-up questions needed? Yes.
Follow-up: [intermediate question 1]
Intermediate answer: [answer after retrieval]
Follow-up: [intermediate question 2]
Intermediate answer: [answer after retrieval]
Final answer: [synthesized response]
Example:
Question: "Who was president when the iPhone was released?"
Are follow-up questions needed? Yes.
Follow-up: When was the iPhone first released?
Intermediate answer: June 29, 2007
Follow-up: Who was the US president in June 2007?
Intermediate answer: George W. Bush
Final answer: George W. Bush was president when the iPhone
was released in June 2007.
When to use it: Factoid chains where each answer feeds the next question. Particularly good for temporal reasoning and entity resolution.
4. Chain-of-Verification (CoVe): Trust but Verify
A different philosophy: generate an answer first, then verify it. This catches hallucinations and improves factual accuracy.
The pattern:
Draft Answer → Generate Verification Questions → Retrieve Evidence → Check Claims → Revise
Implementation:
def cove_rag(query: str):
# Generate initial answer
initial_docs = retriever.search(query)
draft = llm.generate(query, context=initial_docs)
# What should we verify?
verification_questions = llm.generate(f"""
Given this answer: {draft}
What specific facts should be verified?
""")
# Verify each claim
evidence = {}
for question in verification_questions:
evidence[question] = retriever.search(question)
# Revise if needed
return llm.generate(f"""
Original answer: {draft}
Verification evidence: {evidence}
Revise any incorrect claims.
""")
When to use it: High-stakes applications where accuracy matters more than speed. Legal research, medical information, and financial analysis.
Trade-offs: Highest accuracy but also highest latency. Multiple retrieval and generation rounds.
5. FLARE: Retrieve Only When Uncertain
Forward-Looking Active Retrieval (FLARE) is elegant: generate the answer incrementally, but only retrieve when the model's confidence drops.
The insight: Most sentences don't need retrieval. Only fetch when the model is uncertain.
The pattern:
def flare_rag(query: str):
answer = ""
while not complete(answer):
# Generate next sentence with confidence
next_part, confidence = llm.generate_with_confidence(
query, partial=answer
)
if confidence < THRESHOLD:
# Low confidence → retrieve first
docs = retriever.search(next_part)
next_part = llm.generate(query, partial=answer, context=docs)
answer += next_part
return answer
When to use it: Long-form generation where most content is straightforward but some claims need grounding.
Trade-offs: Efficient (fewer retrievals) but requires confidence estimation, which adds implementation complexity.
6. Tree of Thoughts: Explore Multiple Paths
For truly ambiguous questions, a single reasoning path may not be enough. Tree of Thoughts explores multiple approaches and selects the best.
The pattern:
Generate 3 approaches → Pursue each with retrieval → Evaluate → Select best
Example:
Question: "Why did the startup fail?"
Branch 1: Market analysis angle
→ Retrieve market data, competition analysis
→ Conclusion: The Market was saturated
Branch 2: Financial angle
→ Retrieve funding history, burn rate data
→ Conclusion: Ran out of runway
Branch 3: Execution angle
→ Retrieve team changes, product pivots
→ Conclusion: Too many pivots, lost focus
[Evaluate branches]
Best answer: A Combination of factors—saturated market made growth
expensive, which accelerated the burn rate, leading to funding pressure
that caused desperate pivots.
When to use it: Ambiguous or open-ended questions where multiple interpretations are valid.
Trade-offs: Highest quality for complex questions, but expensive (3x+ the compute).
7. Step-Back Prompting: Zoom Out First
Sometimes you need context before specifics. Step-back prompting asks a more general question first.
The pattern:
Original Question → Abstract to General Question → Retrieve General Context → Retrieve Specifics → Combine
Example:
Original: "Why did the 2008 financial crisis hit Iceland so hard?"
Step back: "What makes small economies vulnerable to global
financial crises?"
[Retrieve general principles about small economy vulnerability]
[Retrieve Iceland-specific 2008 crisis data]
[Combine for comprehensive answer]
When to use it: Conceptual questions that benefit from a broader context. "Why" questions often work well with this approach
Choosing the Right Approach
| If your query is... | Use... |
| Predictable, parallelizable | Query Decomposition |
| Complex, multi-hop | ReAct |
| A chain of dependent facts | Self-Ask |
| High-stakes, accuracy-critical | Chain-of-Verification |
| Long-form with occasional facts | FLARE |
| Ambiguous, multiple valid angles | Tree of Thoughts |
| Conceptual, needs context | Step-Back |
In practice, you'll likely combine approaches. Start simple (decomposition), add ReAct for complex queries, and layer in verification for critical applications.
Implementation Tips
1. Set iteration limits. ReAct and similar patterns can loop forever. Cap at 5-7 iterations.
2. Design your action space carefully. Keep it minimal:
search[query] - semantic search
lookup[term] - exact match
answer[response] - terminate
3. Format observations well. Include source attribution so the model can reason about source quality:
def format_results(results):
return "\n".join([
f"[Source {i+1} - {r.metadata['source']}]: {r.text[:500]}"
for i, r in enumerate(results[:3])
])
4. Log everything. The reasoning trace is gold for debugging. Store thoughts, actions, and observations.
5. Handle failures gracefully. When retrieval returns nothing:
if not results:
observation = "No results found. Try a different search angle."
This guides the model to adapt rather than hallucinate.
The Architecture
A production CoT-RAG system has distinct layers:
| Layer | Description |
| Orchestration Layer | Controls flow, manages state |
| Reasoning Layer (LLM) | Thinks, plans, synthesizes |
| Action Layer | Search, lookup, calculate |
| Retrieval Layer | Vector DB, hybrid search |
| Data Layer | Documents, embeddings |
The orchestration layer is key. It parses LLM outputs, routes to actions, manages conversation state, and enforces termination conditions.
Evaluation Matters
How do you know if your CoT-RAG system is working? Track these metrics:
Retrieval quality:
Reasoning quality:
Answer quality:
Build evaluation sets with complex, multi-hop questions. Compare single-shot RAG against your CoT approach. The differences will be stark.
Conclusion
Standard RAG is powerful but brittle. It assumes one retrieval is enough, that you know what to search for upfront, and that the answer exists in a single chunk.
Chain of Thought retrieval breaks these assumptions. It lets your system reason about what it needs, adapt when initial retrievals fall short, and synthesize across multiple sources.
The techniques range from simple (query decomposition) to sophisticated (tree of thoughts). Start simple, measure what breaks, and add complexity where needed.
The goal isn't to implement every technique. It's to build a system that thinks through problems the way a skilled researcher would: methodically, adaptively, and thoroughly.
Your RAG system shouldn't just retrieve. It should be reasonable.
Further reading:
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (Wei et al., 2022)
ReAct: Synergizing Reasoning and Acting in Language Models (Yao et al., 2023)
Self-Ask: Measuring and Improving the Compositional Reasoning of Large Language Models (Press et al., 2022)
Chain-of-Verification Reduces Hallucination in Large Language Models (Dhuliawala et al., 2023)
