Appearance
Phase 4 Interactive Flashcards: Context Engineering & Agentic RAG
Master the architectural concepts of Phase 4 (Context Engineering & Agentic RAG) with interactive flashcards. Test your recall on the Three-Layer Memory Model, RAG Evolution, Indexing Patterns, Question Transformations, RAG Triad Evaluations, and Zero-Trust RAG Security.
📚 Deck 1: The Context Paradigm & Three-Layer Memory Model
Master the 70% error root cause, Working Memory, Episodic Memory, and Semantic Memory:
🗂️ Deck 1: Context Paradigm & Memory Model
Card 1 of 3Context ParadigmClick or press Space to flip ↺
Why do industry benchmarks show that ~70% of production LLM failures stem from context rather than model capability?
Context Paradigm • AnswerClick to flip back ↻
LLMs are probabilistic reasoning engines conditioned on their prompt context. If provided with noisy, truncated, stale, or conflicting context, even state-of-the-art models will produce hallucinations or irrelevant outputs regardless of raw parameter count.
💡 Architect Takeaway: Context engineering (providing the right information at the right time) has a far higher ROI than prompt wording tweaks.
📚 Deck 2: The Evolution of RAG & Agentic Architectures
Master Naive RAG, Advanced RAG (Hybrid + Re-Ranking), Agentic RAG, GraphRAG, and Small-to-Big retrieval:
🗂️ Deck 2: RAG Evolution & Agentic Architectures
Card 1 of 3RAG EvolutionClick or press Space to flip ↺
What distinguishes Naive RAG from Advanced RAG and Agentic RAG?
Naive (Cosine Top-k) → Advanced (Hybrid + Reranking) → Agentic (Self-Correction & Routing)RAG Evolution • AnswerClick to flip back ↻
1. Naive RAG: Fixed chunking, single dense cosine lookup, direct synthesis (brittle, 60-70% accuracy). 2. Advanced RAG: BM25 + Dense Hybrid search, Cross-Encoder Re-Ranking, and Pre/Post-retrieval processing. 3. Agentic RAG: Dynamic query routing, autonomous tool selection, multi-step retrieval loops, and self-reflection.
💡 Architect Takeaway: Agentic RAG converts static one-shot retrieval into an adaptive, self-evaluating workflow.
📚 Deck 3: Indexing Patterns & Context Window Management
Master ParentDocumentRetriever, MultiVectorRetriever, and solving the "Lost in the Middle" attention trough:
🗂️ Deck 3: Indexing Patterns & Context Window
Card 1 of 3Indexing PatternsClick or press Space to flip ↺
How does ParentDocumentRetriever operate and what problem does it solve?
Embed Small Child Chunks → Retrieve & Return Full Parent ChunksIndexing Patterns • AnswerClick to flip back ↻
Large chunks dilute vector specificity, while tiny chunks lose vital surrounding context. ParentDocumentRetriever splits documents into small child chunks for vector indexing, but stores a pointer to the parent chunk in a Docstore. When a child matches, the parent chunk is fed to the LLM.
💡 Architect Takeaway: Solves the chunking dilemma between retrieval granularity and reading comprehension.
📚 Deck 4: Question Transformations
Master Rewrite-Retrieve-Read, Multi-Query with RRF, Step-Back queries, and HyDE:
🗂️ Deck 4: Question Transformations
Card 1 of 3Question TransformationsClick or press Space to flip ↺
What is the Rewrite-Retrieve-Read paradigm and when is it necessary?
Question Transformations • AnswerClick to flip back ↻
User queries are often colloquial, ambiguous, or reference prior conversational pronouns. An LLM rewrites the raw user query into an optimized, self-contained search query before passing it to the vector retriever, significantly boosting retrieval relevance.
💡 Architect Takeaway: Always rewrite ambiguous or conversational user input before executing vector lookups.
📚 Deck 5: The RAG Triad & Automated Offline Evaluations
Master Answer Faithfulness, Context Precision, Answer Relevance via Ragas, and automated CI/CD Pytest gates:
🗂️ Deck 5: RAG Triad & Evaluations
Card 1 of 3RAG EvaluationClick or press Space to flip ↺
What are the three pillars of the RAG Triad in automated evaluation?
Faithfulness | Context Precision | Answer RelevanceRAG Evaluation • AnswerClick to flip back ↻
1. Faithfulness (Groundedness): Is every claim in the generated answer strictly supported by the retrieved context? (Prevents hallucination). 2. Context Precision: Are the retrieved chunks relevant to the query and free of extraneous noise? 3. Answer Relevance: Does the generated answer directly address the original user query?
💡 Architect Takeaway: The RAG Triad isolates whether a failure occurred in the retriever or the generator.
📚 Deck 6: Compulsory Security (Document Provenance & Vec2Text Defense)
Master HMAC/RSA document signatures against Corpus Poisoning (BadRAG) and embedding perturbation against Vec2Text inversion:
🗂️ Deck 6: Compulsory Security
Card 1 of 2Compulsory SecurityClick or press Space to flip ↺
What is Corpus Poisoning (BadRAG) and how does Cryptographic Document Provenance prevent it?
Compulsory Security • AnswerClick to flip back ↻
In BadRAG, an attacker injects subtly poisoned documents into the knowledge base designed to rank #1 for target queries and trigger backdoors or fraudulent answers. Provenance requires every document to carry an HMAC/RSA cryptographic signature verifying its authorized source, author, and timestamp before ingestion.
💡 Architect Takeaway: Never index unauthenticated or unsigned third-party documents into your vector store.