Hybrid search and RAG retrieval that actually returns the right thing
How I design hybrid search and RAG retrieval systems — combining keyword and vector search, reranking, and evaluation so retrieval is precise on identifiers and strong on meaning.
in-depth guides in this pillar
supporting essays in the same cluster
capability lanes this work connects to
Most retrieval systems in production are quietly worse than their demo. Pure vector search understands meaning but fumbles the exact identifiers — names, codes, versions — that real queries hinge on, and teams discover it only after shipping. Hybrid retrieval, fused and reranked, is the engineering response to that gap.
These guides are how I actually build it: combining lexical and semantic retrieval, fusing the rankings deliberately, reranking for real relevance, and — the part most teams skip — evaluating against the queries you actually get, so search is a system you improve on purpose rather than one you ship once and hope holds.
In-depth, code-backed guides.
Metadata Filtering in Vector Search: Pre-Filter, Post-Filter, and the Recall Cliff
How metadata filtering actually behaves in vector search: why post-filtering breaks tenant isolation, when selective filters collapse ANN recall, and what to do.
Embedding Model Selection for Retrieval: How to Choose Without Trusting a Leaderboard
How I pick an embedding model for retrieval: the constraints that decide it before quality does, a bake-off you can run, and the re-index nobody prices.
How to Evaluate RAG Retrieval: Eval Sets, Metrics, and Ship Gates
Retrieval eval is 20% metrics and 80% eval set. How I build one that survives re-indexing, which number to read at which k, and how to gate a change.
RAG Chunking Strategy: How to Split Documents So Retrieval Works
Chunking decides what your retriever can find. Here's how I split real documents — structure-first boundaries, parent-child units, and how to prove it worked.
Why Is My RAG Retrieval Bad? A Diagnostic Order of Operations
Bad RAG retrieval is four or five distinct failures wearing one costume. Here's how I localize which one you have before changing anything.
Reranking in RAG: How a Cross-Encoder Fixes Retrieval Quality
A cross-encoder reranker re-scores your top candidates by reading query and document together. Here's how I add one, size it, and prove it worked.
How to Build Hybrid Search with pgvector and BM25 in Postgres
Build hybrid search in Postgres with pgvector, tsvector, and RRF in one SQL query — the schema, index tuning, and when you actually need real BM25.
Hybrid Search vs Vector Search: Why RAG Retrieval Needs Both
Vector search understands meaning but fumbles exact identifiers; keyword search is the opposite. Here's how I build hybrid retrieval that does both — with fusion, reranking, and a pgvector setup.
Field notes and opinionated takes in the same cluster.
Hybrid Search in Practice: Tuning Relevance Without Lying to Yourself
Relevance tuning is an operational discipline, not a one-time configuration. A deep dive into evaluation metrics, bias suppression, and feedback loops for intelligence systems.
The Hybrid Search Engine: Combining Lexical and Semantic Ranks
OSINT relevance is multi-modal. A technical exploration of why keywords fail and how to fuse BM25 with Vector Embeddings for operator-grade retrieval.
The delivery lanes this work maps to.
Correlation and scoring
Entity resolution, de-duplication, ranking, and confidence models for turning noisy signals into usable intelligence.
Collection and orchestration
Browser automation, distributed workers, scheduling, and fleet-level recovery for public-data systems that need to keep working under drift.
Retrieval returning the wrong things in production?
This is exactly the kind of system I get brought in to design, audit, and make dependable. If that's where you are, let's talk.
New guides in this pillar, by email
I publish an in-depth engineering guide most weeks. Drop your email and I'll send new ones as they land — no noise.