How Does Semantic Chunking Work Under the Hood?
TL;DR
A complete, up-to-date breakdown of under the hood for developers and founders. It covers the core ideas, the trade-offs that matter, a practical workflow, real numbers, and the questions people ask most — written to be skimmed, applied, and shared.
Key takeaways
- Never embed a query with one model and your corpus with another; the query and document vectors must live in the same embedding space.
- Add a cross-encoder reranker over your top candidates; it is one of the highest-leverage, lowest-effort quality wins in a RAG pipeline.
- Reach for GraphRAG when questions require connecting facts across many documents; keep plain vector RAG for direct lookups where it is cheaper and simpler.
- Combine dense semantic search with sparse keyword search (BM25) using hybrid retrieval, because each catches failures the other misses.
- RAG is retrieval plus generation: fix the retrieval half first, because a great model cannot answer from context it never received.
This is a practical, up-to-date guide to Under the Hood — what it is, why it matters in 2026, and how to apply it in real projects. It is written for developers and founders who want clear answers and proven best practices, not filler.
Whether you're just starting out or leveling up, treat this as a working reference you can return to. Every section is built to be skimmed, applied, and shared.
Common failure modes and pitfalls
The most common RAG failures live in retrieval, not the model: if the right chunk is never fetched, no amount of prompt engineering will recover the answer. Frequent culprits include mismatched embedding models for query and corpus, chunking that fragments the answer, missing or wrong metadata filters, and stale indexes that lag behind the source documents. A subtler risk is retrieval poisoning, where malicious or low-quality content in the knowledge base is retrieved and then repeated by the model, since RAG grounds but does not verify. RAG also reduces but does not eliminate hallucination, so answers should be constrained to cite sources and to decline gracefully when the retrieved context does not actually contain the answer.
Semantic versus keyword versus hybrid search
Keyword search, classically BM25, matches on exact terms and excels at precise identifiers, product codes, names, and rare tokens that embeddings can blur together. Semantic search over embeddings captures meaning and paraphrase, so it finds relevant passages even when the wording differs from the query. Each approach fails where the other is strong, which is why hybrid search, running both and fusing the results, is now a common default. A widely used fusion method is Reciprocal Rank Fusion, which combines ranked lists without needing the two systems' scores to be on the same scale, and most mature vector engines now expose hybrid retrieval directly.
GraphRAG and structured retrieval
Plain vector RAG retrieves passages independently, which works for direct lookups but struggles with questions that require synthesizing information scattered across many documents. GraphRAG, introduced by Microsoft Research in 2024, first uses an LLM to extract entities and relationships into a knowledge graph, then clusters and summarizes that graph so retrieval can operate over structured, connected knowledge. This helps with global sensemaking questions like "what are the main themes across this corpus" that flat similarity search answers poorly. The tradeoff is cost and complexity, since building and maintaining the graph consumes many LLM calls, so GraphRAG is best reserved for corpora where cross-document reasoning genuinely matters rather than as a default for every application.
Getting started and where the field is heading
A pragmatic first build is small: a handful of well-chunked documents, a solid off-the-shelf embedding model, pgvector or a lightweight store like Chroma, hybrid search, and a reranker, wired together with a framework such as LlamaIndex or LangChain or with plain code. Prove it works on a real evaluation set before scaling infrastructure, because premature adoption of a distributed vector database often adds complexity without solving the actual retrieval problems. Looking ahead, agentic retrieval that plans multi-step searches, longer context windows that shift some burden away from aggressive chunking, and multimodal embeddings over images and tables are all active areas. The durable lesson is that retrieval quality, evaluation discipline, and clean data pipelines matter more than the specific database, and those fundamentals will outlast any single vendor.
What retrieval-augmented generation actually is
Retrieval-augmented generation, or RAG, is a pattern that grounds a large language model in external data by fetching relevant text at query time and inserting it into the prompt. Instead of relying only on the frozen knowledge baked into the model's weights, the system retrieves passages from a knowledge base and asks the model to answer using that supplied context. The approach was formalized in a 2020 paper from Facebook AI Research and has since become the standard way to make LLMs answer questions about private documents, recent events, or specialized domains. Its appeal is practical: you can update the knowledge base without retraining the model, and you can point to the retrieved passages as evidence for an answer.
Reranking for precision at the top
Retrieval typically returns a few dozen plausible candidates, but the generator can only use a handful, so the ordering of those top results is what actually reaches the model. A reranker is a cross-encoder that reads the query and each candidate passage together and scores their relevance directly, which is far more accurate than the independent vector similarity used during first-stage retrieval. Because cross-encoders are too slow to run over an entire corpus, they are applied only to the shortlist, giving a strong precision boost for modest added latency. Hosted rerankers such as Cohere Rerank and open cross-encoder models from the Sentence-Transformers ecosystem make this one of the easiest high-impact upgrades to a RAG stack.
Under the Hood: Key Facts and Data
According to recent industry research and the official documentation linked below:
- As of 2025, PostgreSQL with the pgvector extension is one of the most popular ways teams add vector search, because it lets them keep vectors, relational data and transactions in a database they already run.
- The MTEB (Massive Text Embedding Benchmark) leaderboard on Hugging Face has become the de facto public scoreboard for comparing embedding models across dozens of retrieval, classification and clustering tasks.
- The HNSW (Hierarchical Navigable Small World) algorithm, published in 2016, is the most widely adopted approximate-nearest-neighbor index and underpins Qdrant, Weaviate, Milvus, pgvector, Elasticsearch and most other vector engines.
Quick-Reference Summary
A map of what this guide covers:
| Topic | What you'll learn |
|---|---|
| Common failure modes and pitfalls | The most common RAG failures live in retrieval |
| Semantic versus keyword versus hybrid search | Keyword search, classically BM25, matches on exact terms and excels at precise identifiers, product codes, names, and |
| GraphRAG and structured retrieval | Plain vector RAG retrieves passages independently |
| Getting started and where the field is heading | A pragmatic first build is small: a handful of well-chunked documents, a solid off-the-shelf embedding model, pgvector |
| What retrieval-augmented generation actually is | Retrieval-augmented generation, or RAG, is a pattern that grounds a large language model in external data by fetching |
| Reranking for precision at the top | Retrieval typically returns a few dozen plausible candidates |
How to Get Started with Under the Hood
A simple path that works:
- Learn the fundamentals of Under the Hood from primary sources, not just tutorials.
- Build one small, real project end to end.
- Get feedback, refactor, and add tests.
- Ship it publicly and document what you learned.
- Repeat with a slightly harder project each time.
Build It with a World-Class Full Stack Developer
Sandeep Kumar Chaudhary is a full stack world-class developer. If you want to turn this into a real, production-ready product, get in touch — message directly on WhatsApp at +9779802348957 for a fast, no-pressure consult.
You can also explore the projects already shipped to thousands of users, or start a conversation here.
Final Thoughts
Never embed a query with one model and your corpus with another; the query and document vectors must live in the same embedding space. The developers and teams who win in 2026 pair strong fundamentals with consistent shipping. Start small, stay curious, build in public, and revisit this guide as your skills grow.
Sources and Further Reading
Frequently Asked Questions
How Does Semantic Chunking Work Under the Hood?
Keyword search, classically BM25, matches on exact terms and excels at precise identifiers, product codes, names, and rare tokens that embeddings can blur together. Semantic search over embeddings captures meaning and paraphrase, so it finds relevant passages even when the wording differs from the query. This guide covers under the hood end to end — core concepts, best practices, concrete data, and a step-by-step approach you can apply right away.
What is a reranker and do I need one?
A reranker is a model, usually a cross-encoder, that reads the query and each candidate passage together and scores their relevance directly, which is more accurate than the independent similarity used during initial vector retrieval. You apply it only to the top candidates from first-stage retrieval, reordering them so the best passages reach the model. It is one of the highest-leverage, lowest-effort quality improvements in a RAG pipeline, so for most applications it is worth adding.
How do I evaluate a RAG system?
Measure retrieval and generation separately, because a good answer needs both. Evaluate retrieval with information-retrieval metrics such as recall at k and mean reciprocal rank against a labeled set of questions with known relevant chunks, and evaluate generation on faithfulness and answer relevance, often with frameworks like RAGAS or an LLM-as-judge. The key discipline is to assemble a representative evaluation set of real questions early so every change can be judged with numbers.
When should I use GraphRAG instead of regular vector RAG?
Use GraphRAG when your questions require connecting facts spread across many documents or summarizing an entire corpus, which flat vector retrieval handles poorly. GraphRAG builds a knowledge graph of entities and relationships and lets retrieval operate over that structure, but it costs many extra LLM calls to construct and maintain. For direct lookups where the answer sits in one or a few passages, plain vector RAG is cheaper, simpler, and usually good enough.
What is hybrid search and why does it help?
Hybrid search runs both keyword search, usually BM25, and semantic vector search, then fuses the two result lists. Keyword search nails exact terms like names, codes, and rare tokens, while semantic search captures meaning and paraphrase, so each covers the other's blind spots. Fusing them, often with Reciprocal Rank Fusion, typically produces more robust retrieval than either method alone, which is why it has become a common default.
Sandeep Kumar Chaudhary
Full Stack Software Developer· Nepal's SEO, AEO, GEO & AIO expert and share-market educator. More about me