📖 8 min read · By ['[“HoangNM105”]']

Quiz & Appendix - Advanced#

Knowledge Summary#

Section 1: Optimization#

Skills Learned#

  • Advanced Indexing (Semantic Chunking, HNSW, Multi-Vector)

  • Hybrid Search (Vector + BM25 + RRF)

  • Query Transformation (HyDE, Decomposition)

  • Post-Retrieval (Reranking, MMR)

Key Takeaways#

  • HNSW dramatically improves search speed

  • Reranking adds precision, MMR adds diversity

Section 2: GraphRAG#

Skills Learned#

  • Graph Database concepts (Nodes, Edges)

  • Knowledge Graph construction

  • GraphRAG architecture

  • Entity/Relation extraction

Key Takeaways#

  • Entity extraction is critical

Section 3: LangGraph & Agentic AI#

Skills Learned#

  • LangGraph foundations (State, Nodes, Edges)

  • Agentic patterns (Reflection, Planning)

  • Tool calling (Tavily Search)

  • Multi-agent collaboration

  • Human-in-the-loop

  • Streaming with FastAPI

Key Takeaways#

  • State management is core

  • Reflection improves quality

  • Planning handles complexity

  • Agents extend LLM capabilities

Section 4: LLMOps#

Skills Learned#

  • RAGAS evaluation metrics

  • LangFuse & LangSmith observability

  • Experimental comparison

  • Production best practices

Key Takeaways#

  • Evaluation is essential

  • Observability enables debugging

  • Each architecture has trade-offs

Multiple Choice Questions#

Part 1: Optimization#

Question 1: How does semantic chunking differ from fixed-size chunking?#

  • A) Splits based on character count

  • B) Splits based on semantic meaning

  • C) Splits randomly

  • D) No difference

Question 2: What does the ‘M’ parameter in HNSW indexing represent?#

  • A) Number of vectors in index

  • B) Number of connections per node

  • C) Embedding dimension

  • D) Number of layers

Question 3: What does Multi-Vector Retrieval store?#

  • A) Only original documents

  • B) Multiple embeddings for each document (summary, questions, etc.)

  • C) Multiple copies of the same embedding

  • D) Only queries

Question 4: What is BM25?#

  • A) Vector similarity algorithm

  • B) Keyword-based ranking algorithm

  • C) Embedding model

  • D) LLM model

Question 5: What is RRF (Reciprocal Rank Fusion) used for?#

  • A) Embed documents

  • B) Generate answers

  • C) Merge rankings from multiple retrievers

  • D) Chunk documents

Question 6: What is the core principle of HyDE (Hypothetical Document Embeddings)?#

  • A) Translating natural language to SQL

  • B) Generating a fake hypothetical answer, then embedding it for search

  • C) Breaking a complex query into sub-questions

  • D) Using keyword search only

Question 7: When is Query Decomposition most effective?#

  • A) For simple, single-fact lookups

  • B) When the query asks about multiple distinct topics or requires multi-step reasoning

  • C) When the user provides keywords only

  • D) When using an image encoder

Question 8: What is a potential risk of using HyDE?#

  • A) It is too fast

  • B) It requires a knowledge graph

  • C) Hallucinated hypothetical answers might lead retrieval astray

  • D) It cannot handle English text

Question 9: Which technique involves generating multiple variations of a user query to improve recall?#

  • A) Multi-Query Retrieval / Expansion

  • B) Vector Quantization

  • C) Semantic Chunking

  • D) Graph Traversal

Question 10: What is “Step-Back Prompting” in the context of Query Transformation?#

  • A) Asking the user to simplify the question

  • B) Generating a more abstract, high-level question to retrieve broad context first

  • C) Ignoring the prompt entirely

  • D) Reversing the order of words in the query

Question 11: How does a Cross-Encoder differ from a Bi-Encoder?#

  • A) Cross-Encoder separates query and document

  • B) Cross-Encoder processes query and document together (Joint extraction)

  • C) Bi-Encoder is slower than Cross-Encoder

  • D) There is no difference

Question 16: What does the Lambda parameter in MMR control?#

  • A) The number of chunks retrieved

  • B) The embedding dimension

  • C) The balance between Relevance vs Diversity

  • D) The learning rate of the model

View Part 1 Answer Key

  • Q1: B

  • Q2: B

  • Q3: B

  • Q4: B

  • Q5: C

  • Q6: B - HyDE bridges the gap between question vectors and answer vectors.

  • Q7: B - Complex questions often need to be solved in parallel or sequence.

  • Q8: C - If the LLM hallucinates facts in the HyDE document, the vector search will look for those incorrect facts.

  • Q9: A - Different phrasings capture different semantic angles of the same intent.

  • Q10: B - Abstraction helps retrieve foundational knowledge that specific details might miss.

  • Q11: B - Cross-Encoder captures deep interaction but is slower; Bi-Encoder is fast but less precise.

  • Q16: C - The trade-off between relevance (λ=1) and diversity (λ=0).

Part 2: GraphRAG#

Question 17: In a graph database, what do Nodes and Edges represent?#

  • A) Nodes = data, Edges = code

  • B) Nodes = entities, Edges = relationships

  • C) Nodes = queries, Edges = answers

  • D) No difference

Question 18: What is Cypher?#

  • A) Encryption algorithm

  • B) Query language for Neo4j

  • C) Embedding model

  • D) Python library

Question 19: What role does “Community Detection” (e.g., Leiden algorithm) play in GraphRAG?#

  • A) It finds users with similar IP addresses

  • B) It groups related nodes to generate high-level summaries/themes

  • C) It detects spam emails

  • D) It encrypts the database

Question 20: What is the primary difficulty in “Entity Extraction” for GraphRAG?#

  • A) Text is too short

  • B) LLMs usually refuse to extract entities

  • C) Entity resolution / disambiguation (e.g., “Apple” vs “Apple Inc.”)

  • D) Storing vectors is expensive

Question 21: Which GraphRAG capability allows finding connections between indirect neighbors (A->B->C)?#

  • A) Vector Similarity

  • B) Semantic Chunking

  • C) Multi-hop Traversal

  • D) Keyword Matching

Question 22: When is GraphRAG better than standard RAG?#

  • A) For simple fact lookups

  • B) When strict low latency is required

  • C) For complex relational queries and multi-hop reasoning

  • D) When costs must be minimized

View Part 2 Answer Key

  • Q17: B

  • Q18: B

  • Q19: B - Community detection identifies clusters of concepts to answer “global” questions.

  • Q20: C - Ensuring the graph is connected correctly requires resolving entity duplicates.

  • Q21: C - Traversal follows edges to find distant connections.

  • Q22: C - When there are many relational queries and multi-hop reasoning requirements.

Part 3: LangGraph & Agents#

Question 23: What is State in LangGraph?#

  • A) Database state

  • B) Shared data between nodes

  • C) API state

  • D) UI state

Question 24: What are conditional edges used for?#

  • A) Routing based on state

  • B) Optimize performance

  • C) Save memory

  • D) Log data

Question 25: What does Andrew Ng’s Reflection pattern do?#

  • A) Speed up inference

  • B) Self-evaluate and iteratively improve output

  • C) Reduce cost

  • D) Compress data

Question 26: What does tool calling allow an LLM to do?#

  • A) Train faster

  • B) Access external data and perform actions

  • C) Generate images

  • D) Compress text

View Part 3 Answer Key

  • Q23: B

  • Q24: A

  • Q25: B

  • Q26: B

Part 4: LLMOps#

Question 27: What does the RAGAS metric “Faithfulness” measure?#

  • A) Answer speed

  • B) Answer grounded in context (no hallucinations)

  • C) Answer length

  • D) Cost

Question 28: What is the difference between Context Precision and Recall?#

  • A) Precision is speed; Recall is accuracy

  • B) Precision is relevant/retrieved; Recall is relevant/total_relevant

  • C) Precision is quantity; Recall is quality

  • D) Precision is for vectors; Recall is for graphs

Question 29: Which RAGAS metric measures if the generated answer actually addresses the user’s question?#

  • A) Faithfulness

  • B) Answer Relevancy

  • C) Context Recall

  • D) Latency

Question 30: What is the main difference between LangFuse and LangSmith?#

  • A) LangFuse is open-source/self-hosted; LangSmith is a managed SaaS with deep LangChain integration

  • B) LangFuse is for images; LangSmith is for text

  • C) LangFuse is paid; LangSmith is free

  • D) LangSmith is only for OpenAI models

Question 31: In Observability, what does “Tracing” refer to?#

  • A) Drawing graphs

  • B) Tracking the step-by-step execution flow of chains/agents

  • C) Finding the IP address of the user

  • D) Copying data to a backup

Question 32: Why is “Sampling” important in Production Observability?#

  • A) To save storage costs and reduce noise by not logging 100% of traces

  • B) To test new models

  • C) To improve latency

  • D) To satisfy GDPR

Question 33: Which tool provides an “Edit and Re-run” playground for debugging failed traces?#

  • A) LangFuse

  • B) LangSmith

  • C) Prometheus

  • D) Grafana

Question 34: When should you use Hybrid RAG (Graph + Advanced)?#

  • A) When you have no budget

  • B) When latency is the top priority

  • C) When you need the absolute best quality for mixed query types

  • D) When you only have unstructured text with no entities

View Part 4 Answer Key

  • Q27: B

  • Q28: B - Precision: Relevant/Retrieved (Noise level); Recall: Relevant/Total (Coverage).

  • Q29: B - Answer Relevancy measures alignment with the query, not truthfulness.

  • Q30: A - LangFuse is OS friendly; LangSmith is SaaS.

  • Q31: B - Tracing visualizes the black box execution.

  • Q32: A - Monitoring every single request in high-traffic apps is expensive.

  • Q33: B - LangSmith’s deep integration allows re-running traces with modified prompts.

  • Q34: C - When best quality is needed, there are mixed query types, and budget is not an issue.

Appendix#

Graph Databases#

Evaluation#

Observability#

GitHub Repositories#

Example Projects#

On this page