#meilisearch · 8sync News

RAG for medical data: improving healthcare AI accuracy

RAG for medical data connects large language models to trusted healthcare datasets like clinical guidelines, PubMed studies, and electronic health records to reduce AI hallucinations and improve accuracy. The guide covers how MedRAG works, the five-step pipeline (ingestion, embedding, indexing, retrieval, generation), real-world use cases including clinical decision support, medical research search, patient chatbots, and drug lookup, plus how to build and evaluate medical RAG systems. Key challenges include HIPAA compliance, data privacy, dataset bias, and integration complexity. Meilisearch is presented as a tool to simplify hybrid search and document indexing within these pipelines.

Query rewriting for RAG: how to improve retrieval accuracy

Query rewriting in RAG pipelines transforms user queries before retrieval to better match the language of knowledge bases and datasets. The guide covers why retrieval quality determines LLM response accuracy, the full workflow from query normalization through generation, and five core techniques: query expansion, decomposition, paraphrasing, multi-query generation, and step-back prompting. Practical Python code examples show how to normalize queries, call an LLM to rewrite them, retrieve documents via vector search, and generate grounded answers. Evaluation metrics (Recall@k, MRR, NDCG) and comparisons with reranking and query expansion are also covered, along with limitations like semantic drift, added latency, and over-expansion.

What is context distillation in AI & how does it improve LLM efficiency?

Context distillation is the process of selecting, filtering, and compressing input data so only the most relevant information is passed to an LLM's context window. It reduces token usage, lowers costs, improves latency, and produces more consistent outputs in enterprise AI workflows. The post covers how it works step-by-step, common techniques (prompt-based supervision, synthetic dataset generation, on-policy distillation, iterative refinement), real-world examples, and how it differs from model distillation and fine-tuning. Meilisearch is presented as a retrieval layer that supports context distillation pipelines by converting documents to vector embeddings and feeding relevant chunks to LLMs.

CarJager: Millions of Vehicle Searches on Meilisearch

CarJager, a French classic car marketplace, migrated its site search from Algolia to Meilisearch Cloud to overcome Algolia's 1,000-entity sorting replica limit, eliminate DevOps overhead, and reduce costs. With ~1.3M searches/month against a ~2,000-listing catalog, Algolia's pricing was disproportionate. The migration was phased — starting with blog search, then core vehicle listings — and completed in days thanks to API familiarity. Results include native sorting across the full catalog, autoscaling that handles traffic peaks without manual intervention, and a significant cost reduction after right-sizing from an L to M instance.

A practical guide to search relevance metrics and evaluation

A comprehensive guide to search relevance metrics covering precision, recall, F1 score, MAP, MRR, nDCG, ERR, AUC, and CTR. Explains formulas and when to use each metric, how to measure relevance with and without manual labeling, how to evaluate search quality in production using logging, A/B testing, and KPIs, and common evaluation mistakes to avoid. Concludes with guidance on choosing the right metrics for different use cases and a brief mention of Meilisearch's ranking capabilities.