Vectorless RAG
A retrieval pattern where the human stays in the loop because the loop is where learning happens.
A framework and technique I apply across case studies, journals, and systems I build.
Standard RAG embeds documents into vector stores and lets the model find relevant chunks. Vectorless RAG inverts this: the human selects, primes, and structures the input. The LLM narrates, not navigates. The retrieval act builds understanding, which is why the human can't be removed without destroying the value.
Applications: Enterprise Software
Context. Designing canonical query patterns for a large legacy relational database. Needed to prove that declared relationships could replace vector search for AI traversal, without hallucinating connections that weren't explicitly configured.
The move. Structured the declared relationship graph as a prompt constraint rather than an embedding. Gave the model one instruction: "traverse declared relationships only, do not infer connections that aren't in this config." Iterated until the query patterns were deterministic.
What would have been lost. If the LLM retrieved on its own, it would have hallucinated relationships that weren't configured. The point was that declared beats inferred: letting the model guess would have undermined the entire thesis.
Context. Building an AI Adoption Governance Framework for regulated enterprise customers like government agencies and health systems. Needed to define where AI can and cannot act autonomously.
The move. Synthesized patterns from customer-facing conversations across sales, support, and governance contexts. Asked: "What are these customers actually afraid of?" Then validated the three governance principles against public regulatory frameworks including FedRAMP and HIPAA.
What would have been lost. An LLM generating governance principles from training data would produce generic AI ethics boilerplate. The specificity (vendor abdication, dark by default, human authority) came from real customer fear, not abstract principle.
Applications: Inyeon AI
Retrieval as signal. Context. Designing CQL v3.0 for Inyeon's journaling app. The question was what counts as structured data: the text of an entry, the tags around it, or the user's choice to write it down at all.
The move. Treated journaling choices as the data model. Which entries got captured, which paradigms got invoked, which moments got flagged OFNR: those selections weren't metadata on top of content, they were the content. Primed Claude with the user's selection history and asked it to reason over what the choices themselves implied.
What would have been lost. If retrieval had been automated, CQL would have collapsed into generic NLP. The paradigm assignments earn their specificity because the retrieval step is a design decision, not an algorithm: the act of choosing is the schema.
Applications: Personal Infrastructure
Context. Building ORÍ Central's nine-database architecture across Notion. Needed deterministic validation that system prompts route correctly across multiple LLM instances.
The move. Created the 42 Test: a fixed prompt sent to every new Claude instance to verify it can read, write, and route to the correct Notion databases. I retrieved the expected schema, fed it alongside the test prompt, and compared actual vs. expected output. The human is the test harness.
What would have been lost. Automated testing of LLM instruction-following doesn't exist in a meaningful way yet. The human performing the retrieval IS the validation layer. You can't automate judgment about whether a system prompt "worked."
Context. Synthetic UX testing for Clause Ink's consent interface. No users yet, needed to validate whether the opt-out toggle was cognitively clear before building it.
The move. Screenshotted the UI mockup. Primed Claude with Lupton's "Design Is Storytelling" table of contents, specifically the chapters on cognitive load and decision architecture. Asked: "Where does this interface create ambiguity about what the user is consenting to?" Then iterated the design based on the critique.
What would have been lost. Without the UX book priming, the critique would have been surface-level ("the button should be bigger"). The book retrieval gave the LLM a vocabulary for cognitive load that it doesn't reliably have on its own.
Why this matters
Every role on my portfolio shares a thesis: declared relationships beat inferred ones. Vectorless RAG is that thesis applied to how I use AI itself. I don't ask models to find what's relevant. I bring what's relevant and ask models to reason over it.
This isn't anti-automation. It's pro-judgment. The retrieval step is where expertise lives: knowing which screenshot matters, which book chapter applies, which schema field is load-bearing. Automating that step doesn't save time. It removes the human from the part of the process where they add the most value.