Few-Shot Prompting with Graph Samples: Learning the Grammar

An LLM is a generalist. It knows how to write a poem or a summary, but it doesn't intuitively know how to "Reason over a Directed Acyclic Graph." If you just give it context and a question, it might miss the subtle logic of a "Reverse Edge" or a "Temporal Constraint." To fix this, we use Few-Shot Prompting. We give the AI 2-3 examples of a Question -> Graph Subgraph -> Correct Answer chain.

In this lesson, we will look at how to build a Graph-Specific Few-Shot Library. We will learn which types of examples are most effective for teaching "Multi-hop Reasoning" and how to use these examples to guide the AI toward a structured, traceable response.

---

1. Why Zero-Shot Fails for Graphs

Zero-Shot Prompt: "Here is the graph. Answer the question." AI Failure: The AI might see (A)-[:OWNS]->(B) but say "B is the boss of A" because it's used to seeing power relationships in text, ignoring the direction of the arrow.

Few-Shot Fix: We provide an example where a similar "Arrow Conflict" is resolved correctly.

• Example 1: (John)-[:OWNS]->(Company).
• Check: Who is the owner?
• Answer: John.

---

2. The Anatomy of a Graph Few-Shot Example

1. The Question: Simple and clear.
2. The Graph Evidence: A small set of 3-5 triplets.
3. The Reasoning Path: A "Chain of Thought" explaining the walk.
4. The Final Answer: Concise and grounded.

Pro Tip: Always include an example where the answer is Unknown. This teaches the AI not to "Hallucinate" a bridge where one doesn't exist in the graph.

---

3. Dynamic Few-Shot Selection

Don't use the same 3 examples for every query.

• If the user asks a Temporal question, retrieve 3 few-shot examples that involve "Dates" and "Sequence."
• If the user asks a Path question, retrieve examples showing "Multi-hop" reasoning.

This is called In-Context Learning (ICL) optimization.

graph TD
    User[Query: How many?] --> Classifier[Intent: Counting]
    Classifier --> Library[Few-Shot Library]
    Library -->|Fetch| E1[Count Example]
    Library -->|Fetch| E2[Count Example 2]
    E1 & E2 --> LLM[Final System Prompt]
    
    style E1 fill:#34A853,color:#fff
    style LLM fill:#4285F4,color:#fff

---

4. Implementation: A Graph Few-Shot Template

EXAMPLE 1:
QUESTION: Who is the senior leader for the Marketing project?
GRPAH CONTEXT:
(Sudeep)-[:ROLE]->(Lead)
(Sudeep)-[:WORKS_ON]->(Marketing)
REASONING: The graph shows Sudeep works on Marketing and his role is Lead.
ANSWER: Sudeep is the senior leader.

---
// REAL QUERY //
QUESTION: {user_query}
GRAPH CONTEXT:
{retrieved_graph}
REASONING:
ANSWER:

---

5. Summary and Exercises

Few-shot prompting is the "Instruction Manual" for the AI's brain.

• Directionality is the most common error in zero-shot graph RAG.
• Negative examples ("I don't know") prevent hallucinations.
• Chain of Thought (CoT) in examples teaches the AI how to "walk."
• Dynamic selection ensures the examples are relevant to the user's current logic puzzle.

Exercises

1. Example Writing: Write a few-shot example that teaches an AI to find the "Founder" of a company using a 2-hop path: (Person)-[:INVESTED_IN]->(Company) AND (Person)-[:AUTHORED]->(Patent).
2. The "Hallucination" Check: If you add an example where the AI says "The graph does not provide enough evidence," how does that change the bot's behavior for vague questions?
3. Visualization: Draw a "Library" of 10 examples. Tag each with a "Topic" (e.g., Dates, People, Sums).

In the final lesson of this module, we will look at data formats: Structured Context: YAML vs Markdown vs JSON for Graphs.