RAG pipeline returning wrong results.Fixed this session.

PM assigned in 10 minutes. RAG engineer starts diagnosis within the hour. Irrelevant results in RAG pipelines are almost always caused by one of four issues: the chunking strategy is wrong chunks are too large (diluting the relevant content with surrounding noise) or too small (splitting context that needs to stay together), the embedding model is not a good semantic fit for your domain (a general-purpose embedding model performs poorly on technical or legal text), the similarity search threshold is too permissive (retrieving chunks that have surface-level word overlap but are semantically unrelated to the query), or the re-ranking step is missing or misconfigured. The engineer diagnoses using evaluation metrics retrieval precision, recall, and RAGAS scores on your real queries and implements targeted fixes.

Yes. Embedding strategy redesign is a Full Day session. This covers: evaluating which embedding model best fits your document domain (OpenAI text-embedding-3-small vs. large, Cohere embed-v3, or a domain-specific fine-tuned model like a legal or medical embedding model), chunking strategy redesign (fixed-size, recursive character splitting, semantic chunking by sentence boundary, or hierarchical chunking with parent-child relationships for preserving context), overlap configuration between chunks to prevent context gaps at boundaries, and metadata enrichment (adding document title, date, section heading, and source URL as filterable metadata so retrieval can be constrained beyond pure semantic similarity). The engineer measures before-and-after retrieval accuracy on your evaluation set before the session closes.

Yes. Engineers are matched to your specific vector database. For Pinecone: namespace management, index configuration (dimension, metric, pod type), metadata filtering with sparse-dense hybrid search, and cost optimisation (serverless vs. pod-based indexing). For Weaviate: schema design with cross-references, hybrid search (BM25 plus vector), generative search module configuration, and multi-tenancy setup. For Qdrant: collection configuration, payload indexing for filtered search, named vector support for multi-modal retrieval, and on-premise deployment configuration. For Chroma: embedding function configuration, persistent storage setup, and collection management. For Redis VSS: hash vs. JSON document storage, HNSW index parameters, and integration with LangChain or LlamaIndex retrieval chains.

Measurement is built into the session deliverable you should never take a RAG fix on faith. The engineer evaluates using RAGAS (Retrieval Augmented Generation Assessment Score), which measures four dimensions: faithfulness (does the answer only make claims supported by the retrieved context?), answer relevancy (does the answer actually address the question asked?), context precision (are the retrieved chunks genuinely relevant to the query?), and context recall (did retrieval find all the information needed to answer correctly?). Before-and-after RAGAS scores are calculated on your real query set. The engineer also builds a test set of ground-truth question-answer pairs from your documents so you can run the evaluation automatically after future changes to the pipeline to catch regressions.

The five most common root causes we diagnose across RAG pipeline engagements: first, chunks that are too large (512-1024 tokens) that contain both the relevant sentence and many surrounding irrelevant sentences the vector representation averages across all content and loses the specific signal; second, the wrong similarity threshold allowing chunks with only superficial keyword overlap to pass the retrieval filter; third, no re-ranking step to re-score and re-order retrieved chunks by relevance before they are sent to the LLM top-k cosine similarity alone is a weak relevance signal; fourth, missing query transformation (the user's question should be rewritten into a retrieval-optimised form before embedding it); fifth, the embedding model dimension mismatch between query and document embeddings from different model versions.