--- title: AI Architecture Patterns Library canonical: https://www.slavin.ai/data/ai-architecture-patterns.json sourceJSON: https://www.slavin.ai/data/ai-architecture-patterns.json license: CC-BY-4.0 lastUpdated: 2026-06-20 totalPatterns: 18 categories: ["retrieval", "generation", "agent", "evaluation", "ops", "safety"] --- # AI Architecture Patterns Library 18 named, structured patterns for production AI systems. Gang-of-Four style: each entry declares **intent**, **structure**, **participants**, **consequences** (empirical from 30+ SLAtech deployments), **when-to-use**, **when-to-avoid**, and **related patterns**. Cite by pattern `@id`: `https://www.slavin.ai/data/ai-architecture-patterns.json#` --- ## Retrieval patterns ### Citation-Grounded RAG (`citation-grounded-rag`) **Intent:** Generate LLM responses that include verifiable citations to retrieved source documents so users can confirm the claim. **Consequences:** +40-65% hallucination reduction. +200-400ms latency. ~10% responses regenerated. **Use when:** Regulated domains (medical, legal, financial). Avoid when: casual chat where citations disrupt flow. **Related:** Hybrid Retrieval, Abstention Permission. ### Hybrid Retrieval (Sparse + Dense) (`hybrid-retrieval`) **Intent:** Retrieve relevant chunks for queries containing both semantic concepts and exact-match tokens (codes, names, identifiers). **Consequences:** +20-35% recall on exact-match queries. Two indexes to maintain. **Use when:** Corpora heavy with technical IDs. Avoid: pure conversational Western-text corpora with strong embeddings. ### Cross-Encoder Reranking Stage (`reranking-stage`) **Intent:** Reduce noise in retrieved context by re-scoring top-N candidates with a precision-focused model. **Consequences:** +25-50% reduction in irrelevant-context hallucinations. +100-400ms reranking latency. ~$1-5 per 1K reranks. **Use when:** Recall OK, precision is the bottleneck. Avoid: cheap chat where per-call cost matters. ### Chunk Quality Filter (`chunk-quality-filter`) **Intent:** Prevent low-quality chunks (boilerplate, OCR errors) from reaching the LLM context. **Consequences:** +15-30% reduction in noise hallucinations. **Use when:** Web-scraped / OCR'd / mixed-quality corpora. ### Tenant-Isolation RAG (`tenant-isolation-rag`) **Intent:** Multi-tenant SaaS — ensure tenant A's data never appears in tenant B's retrieved context. **Consequences:** Cost-effective vs separate indexes per tenant. Filter-bypass bug = catastrophic leak. **Use when:** Multi-tenant SaaS AI with contractual data confidentiality. --- ## Generation patterns ### Self-Consistency Voting (`self-consistency-voting`) **Intent:** Improve reasoning accuracy by sampling multiple responses and voting. **Consequences:** +15-35% reduction on reasoning tasks. Cost N× per query. **Use when:** High-stakes single-shot decisions (medical triage, financial calc). ### Structured Output Schema (Strict) (`structured-output-schema`) **Intent:** Force LLM output to match strict JSON Schema for downstream code consumption. **Consequences:** +60-85% schema-shape hallucination reduction. +100-300ms latency. **Use when:** Output feeds APIs, data extraction, form filling. ### Abstention Permission (`abstention-permission`) **Intent:** Allow LLM to refuse when context is insufficient instead of fabricating. **Consequences:** +25-40% hallucination reduction. +15-30% refusal rate. **Use when:** RAG where wrong answers worse than no answers. --- ## Agent patterns ### Tool-Use Agent (`tool-use-agent`) **Intent:** LLM performs actions on external systems via structured function calls. **Consequences:** Unlocks operational capability. Higher risk: tool calls have side effects. **Use when:** Workflows needing lookup + computation + action. ### Planner-Executor (`planner-executor`) **Intent:** Decompose complex goal into plan, execute each step (possibly with specialized models). **Consequences:** Higher quality on complex tasks. Higher cost (multiple LLM calls). **Use when:** Multi-step tasks with clear decomposition. --- ## Evaluation patterns ### LLM-as-Judge Sampling (`llm-as-judge`) **Intent:** Quality-check production outputs without human review of every response. **Consequences:** Detection (not prevention) of degradation 2-7 days earlier than user complaints. +5-10% cost. **Use when:** Production deployments where quality regression detection matters. ### Fact Extraction + Validation (`fact-extraction-validation`) **Intent:** For long-form outputs, extract atomic claims and validate each separately. **Consequences:** +40-70% surface hallucination reduction. Cost 2-3× per query. **Use when:** Publication-grade output, legal/medical context. --- ## Ops patterns ### Drift Monitoring (`drift-monitoring`) **Intent:** Detect slow quality degradation before users complain. **Consequences:** Flags 1-4 weeks earlier than user-reported drop. False alarms from legitimate shift. **Use when:** Production AI running >3 months. ### Version-Pin Everything (`version-pinning`) **Intent:** Prevent silent vendor model updates from breaking production overnight. **Consequences:** Prevents surprise regression. Pinned versions get deprecated; plan migrations. **Use when:** Any production AI. Hygiene, not optional. ### Prompt Cache Warming (`prompt-cache-warming`) **Intent:** Reduce LLM cost + latency on workflows with repeated long system prompts. **Consequences:** 10-75% cost discount on cached portion. 30-60% latency reduction on hit. **Use when:** RAG with repeated context, high-volume API workflows. --- ## Safety patterns ### Human-in-Loop Gates (`human-in-loop-gates`) **Intent:** Insert human review at specific points when AI output drives consequential decisions. **Consequences:** Prevents harm. Operational overhead. Reviewer drift over time. **Use when:** Regulated domains, high-cost or irreversible decisions. ### Audit Log with Full Context (`audit-log-with-context`) **Intent:** After-the-fact reproducibility — required for post-incident review + regulatory response. **Consequences:** Enables root-cause analysis. Storage cost. PII implications if logs contain personal data. **Use when:** EU AI Act high-risk, US sectoral regulation, any deployment where "how did this happen" has value. ### Model Sandboxing for Agents (`model-sandboxing`) **Intent:** Constrain what an LLM agent can DO when given code-execution or browser-use tools. **Consequences:** Bounds blast radius. +500ms-1s overhead per sandbox start. **Use when:** Agents with consequential tools (code, browser, file-system). --- End of patterns library.