Generative AI Engineering
Learn the ingestion, indexing, retrieval, ranking, grounding, citation, and RAG prompt patterns used to build verifiable AI answers from trusted documents.
19 matching terms
Overview
Retrieval-augmented generation (RAG)
A pattern that retrieves relevant external information and supplies it to a model before generation.
Use it when answers should rely on current, private, or domain-specific documents.
Retrieve policy passages for the question, then answer using only those passages.Ingestion
Chunking
Splitting documents into smaller units that can be indexed and retrieved.
Choose chunk boundaries that preserve meaning while keeping retrieval focused.
Split the handbook by section and keep each heading with its paragraphs.Indexing
Embedding
A numerical vector representing semantic characteristics of text or other content.
Use embeddings to compare semantic similarity beyond exact keyword matches.
Create embeddings for each document chunk and for the user's query.Indexing
Vector store
A system that stores embeddings and supports nearest-neighbor retrieval.
Use it to manage semantic indexes and associated document metadata.
Store each chunk vector with its source URL, title, and access scope.Retrieval
Similarity search
Retrieval based on how close vector representations are under a chosen distance measure.
Use it to find passages with meaning related to the query.
Search the vector store for passages nearest to the query embedding.Retrieval
Retriever
The component that selects candidate documents or chunks for a query.
Configure it to combine query rewriting, filters, search methods, and result limits.
The retriever returns policy chunks the current user is allowed to access.Retrieval
Top-k
The number of highest-ranked retrieval results selected for the next stage.
Tune it to balance evidence coverage against noise and context usage.
Retrieve top-k 8 candidates, then rerank and keep the best 4.Retrieval
Metadata filter
A constraint that narrows retrieval by attributes such as date, product, language, or permissions.
Use it before semantic ranking to enforce scope and access control.
Filter to Korean documents published after 2025 that the user may view.Retrieval
Hybrid search
A search method that combines lexical keyword signals with semantic vector signals.
Use it when exact identifiers and conceptual similarity both matter.
Combine BM25 keyword results with embedding similarity for error code AB-104.Ranking
Reranking
A second-stage model or rule that reorders retrieved candidates for relevance.
Use it to improve precision after collecting a broader candidate set.
Rerank 20 candidate chunks and send the best 5 to the generator.Grounded response
Grounding
Connecting a generated response to supplied or retrieved evidence.
Use it to reduce unsupported claims and make the answer auditable.
Answer only claims supported by the retrieved policy excerpts.Grounded response
Citation
A reference that identifies the source supporting a claim.
Attach citations at the claim level so readers can verify evidence quickly.
A citation can still point to irrelevant evidence; verify that it actually supports the claim.
Include the document title, section, and direct source link for each policy claim.RAG prompt patterns
Query rewriting prompt
An instruction that transforms a user's request into one or more queries better suited to keyword, semantic, hybrid, or filtered retrieval.
Use it for vague, conversational, underspecified, or domain-mismatched questions while preserving the original intent.
Do not silently change the user's intent, add unsupported facts, or weaken access filters.
Rewrite the question for hybrid search. Preserve the original intent, add likely document terminology, and return the original plus up to three rewritten queries.RAG prompt patterns
Grounded-answer instruction
A prompt rule requiring factual claims to be supported by retrieved passages supplied to the generation step.
Use it when answers must be auditable and unsupported model knowledge should not be presented as evidence.
Answer the question using the supplied passages. Attach the supporting source ID to every factual claim and separate any clearly labeled inference.RAG prompt patterns
Context-only answering
A strict generation rule that limits the answer to information explicitly supported by the supplied retrieval context.
Use it for policies, contracts, regulated content, or other tasks where outside model knowledge must not fill gaps.
This rule cannot compensate for poor, incomplete, outdated, or unauthorized retrieval results.
Use only <retrieved_context>. If the context does not contain the answer, return insufficient_evidence instead of using prior knowledge.RAG prompt patterns
No-evidence response and abstention
A defined fallback response used when retrieved evidence is absent, insufficient, contradictory, or below a support threshold.
Use it to prevent confident guessing and to route the case to clarification, another retrieval attempt, or human review.
If no passage directly supports the answer, return status: insufficient_evidence, explain what is missing, and suggest one clarifying question.RAG prompt patterns
Citation format instruction
A prompt contract defining where and how source identifiers must appear in a generated answer.
Use it to produce claim-level references that applications and reviewers can validate consistently.
Formatting a citation does not prove that the cited passage supports the claim; validate citation correctness separately.
After each supported sentence, cite one or more retrieved IDs in [S1] format. Never invent an ID and do not cite a source that does not support the sentence.RAG prompt patterns
Retrieved-context synthesis
Combining complementary evidence from multiple retrieved passages into a coherent answer without erasing source boundaries.
Use it for comparison, summary, or multi-part questions that require more than one passage.
Group passages by subquestion, remove duplicates, preserve dates and scope, then synthesize the answer with claim-level source IDs.RAG prompt patterns
Conflicting-source handling
A prompt rule for detecting and reporting contradictions among retrieved sources instead of silently choosing one.
Use it when documents can differ by version, date, jurisdiction, authority, or interpretation.
Recency alone does not always determine authority; encode the domain's source-priority policy outside or alongside the prompt.
If sources conflict, list each claim with its source, version, and date; identify the conflict; apply an explicit authority rule or request review.Retrieval quality and generation instructions must be designed and evaluated together; a prompt cannot repair missing or unauthorized evidence.
| Stage | Prompt responsibility |
|---|---|
| Before retrieval | Preserve intent while rewriting, expanding, filtering, or decomposing the query |
| Context assembly | Label source IDs, versions, dates, permissions, and passage boundaries |
| Answer generation | Limit claims to evidence, define citation format, and specify abstention behavior |
| After generation | Validate claim support, citation correctness, conflicts, completeness, and access compliance |
No. It reduces the allowed evidence scope, but the answer can still misread passages, combine incompatible sources, attach incorrect citations, or rely on poor retrieval. Validate retrieval recall, claim support, citation correctness, and abstention behavior.
Abstain when no authorized passage directly supports the required claim, evidence is below the defined threshold, sources conflict without an authority rule, or the question is too ambiguous to retrieve safely. Return a structured reason and the next useful action.
Yes when users or downstream systems need to verify claims, but the application must preserve real source identifiers and validate that each cited passage supports the nearby claim. The model must never invent source IDs.
Official architecture guidance on query augmentation, decomposition, rewriting, retrieval, reranking, and accumulated context.
Official documentation for orchestration and generation prompt templates, runtime placeholders, retrieved context, queries, and citation output instructions.
Official guidance for measuring whether generated claims are supported by supplied facts and identifying supporting or contradictory evidence.
Review instructions, constraints, output schemas, task decomposition, reasoning, and verifier patterns.
Budget, assemble, order, compress, cache, and trace retrieved evidence and conversation state.
Measure groundedness, citation quality, completeness, regression, latency, and cost.
Protect retrieval pipelines from indirect prompt injection, unauthorized data access, and unsafe outputs.