Context Engineering
Giving Your AI the Right
Information at the Right Time

The embedding model is not a commodity choice — different models encode different semantic relationships, and the right choice depends on your domain, language patterns, and retrieval task. Vervelo evaluates embedding models (OpenAI text-embedding-3, Cohere Embed, BGE, E5, domain-specific biomedical embeddings) against your actual query distribution and document corpus before selection. For healthcare applications, we evaluate models specifically on clinical terminology, medical abbreviations, and the semantic relationships that matter in clinical contexts — not just general benchmark scores.

Pure dense vector retrieval misses exact keyword matches; pure sparse retrieval misses semantic similarity. Hybrid retrieval combines both — using dense embeddings for semantic relevance and BM25 or TF-IDF for keyword precision — then applies a reranking model (Cohere Rerank, cross-encoder) to score retrieved passages by relevance to the specific query before injection into the context window. Vervelo designs the full hybrid retrieval pipeline including the fusion strategy (Reciprocal Rank Fusion, weighted combination), reranker configuration, and the top-k selection logic that determines how many passages enter the context window.

Vector database selection depends on scale, latency requirements, update frequency, and infrastructure constraints. Vervelo evaluates and configures the right vector store for your use case — Pinecone for fully managed scale, Weaviate for hybrid search with rich filtering, Qdrant for high-performance self-hosted deployments, pgvector for teams that want to stay in Postgres, or Chroma for development and smaller production use cases. Index design decisions — distance metric selection, HNSW parameter tuning, payload indexing for metadata filtering — are benchmarked against your production query patterns before go-live.

Semantic similarity alone is not sufficient for production RAG systems. Retrieved passages must also satisfy structural constraints — document date, source type, author, clinical specialty, payer, geography, or access permission level. Vervelo designs the metadata schema attached to every indexed chunk, and builds the filtered retrieval logic that combines vector similarity with hard constraints. For healthcare applications this includes access control filtering (only retrieve documents the requesting user is authorized to see), temporal filtering (prefer newer clinical guidelines over outdated ones), and source authority filtering (weight content from authoritative clinical sources more heavily).

Many healthcare AI applications require the model to remember facts about a specific patient or user across multiple sessions — clinical history, care preferences, previous interactions, open care gaps, current medications. Vervelo builds long-term memory systems that extract, store, and retrieve user-specific facts: structured memory stores (patient profile databases with FHIR alignment), vector-indexed episodic memory (past interaction summaries retrievable by semantic similarity to the current query), and rule-based memory that surfaces critical facts (allergies, contraindications, active care plans) regardless of semantic relevance. All long-term patient memory is built under HIPAA-compliant data handling and access control requirements.

Beyond session and user-specific memory, AI systems need access to domain knowledge that is not in the model's weights — your organization's specific protocols, formulary policies, network rules, and clinical guidelines. Vervelo designs the semantic memory layer that makes this knowledge retrievable: structured knowledge graphs for relational clinical information, vector-indexed document stores for unstructured guideline content, and hybrid retrieval that combines both for queries that require reasoning across structured and unstructured knowledge simultaneously. For healthcare, this often includes real-time integration with clinical decision support databases, formulary systems, and payer policy repositories.

When raw context exceeds the budget, compression must preserve the information the model needs without introducing distortion. Vervelo builds context compression pipelines using extractive summarization (selecting the most important sentences from retrieved passages), abstractive summarization (using a smaller model to condense long documents into dense summaries before injection), and selective truncation with preservation rules (always keep the most recent N turns, always include critical patient facts regardless of compression ratio). Compression pipelines are evaluated for information retention — measuring whether compressed context produces equivalent model outputs to full context on a representative query set.

Context assembly has a latency cost — retrieval, reranking, compression, and context formatting all add time before the model call begins. Vervelo optimizes the full context assembly pipeline for production latency targets: semantic caching of retrieved passage sets for similar queries (avoiding redundant vector search), prefix caching to reuse common system prompt prefixes across requests, asynchronous pre-fetching of likely-needed context based on conversation trajectory, and tiered retrieval that returns fast approximate results first and enriches with more precise retrieval as the conversation continues. Every optimization is benchmarked against both latency and quality targets to verify the tradeoff is acceptable.