Every choice defensible
Embedding model, retrieval strategy, reranker, decoding parameters: each is a decision the team can defend from first principles, not a default nobody questioned.
← The K-Framework
★ Foundations · Systems ThinkingLayer 3 of 16Visual guide
PILLAR A · LAYER 03 · A.03Algorithmic Fundamentals.
Understand core algorithms.
What a CEO/CTO needs to know
When the system degrades, can your team say whether it was retrieval, rerank, prompt, or model? If the LLM is a black box to them, every fix is a guess.
A retrieval pipeline where every stage is inspectable, so a quality drop has an address.
[WHAT IT IS]
The engineer’s view, in plain language.
We do not import and pray. Knowing why an algorithm works, and when it does not, is non-negotiable. The team that built the system can debug it at the algorithmic level when the abstraction leaks, because abstractions always leak.
[HOW WE BUILD IT]
What “done right” looks like.
Trade-offs in writing
Why this approach and not the obvious alternative is recorded in an ADR, so the reasoning survives the engineer who made it.
Debuggable to the layer
When quality drops, the team isolates the failing stage instead of swapping the whole pipeline and hoping.
[MATURITY LADDER]
Where does your build sit?
Four rungs from absent to production-grade. Level 3 is the target, and the only one that survives a real production incident.
L0
Absent
The pipeline is a black box. Tuning is copy-pasted from blog posts.
L1
Ad-hoc
The team knows the parts but cannot explain why each was chosen.
L2
Managed
Most choices are documented, but debugging still means swapping whole stages.
L3Target
Production-grade
Every algorithmic choice is defensible from first principles and recorded in ADRs; failures isolate to a stage.
[VALIDATE IT YOURSELF]
How to check it’s really there.
You do not need to read the code. Ask these questions and demand these artifacts. Vague answers are the finding.
★ Ask your team
- ?Why did we pick this embedding model and reranker over the alternatives?
- ?When answer quality drops, how do we know which stage caused it?
- ?Who on our team can defend the decoding parameters?
★ Demand to see
- ADRs for embedding, retrieval, rerank, and decoding choices
- A way to evaluate each pipeline stage in isolation
- An engineer who can whiteboard the pipeline end to end
● WHAT L0 LOOKS LIKE
The failure mode, in production.
Treating the LLM as a black box. Stack-Overflow-driven retrieval tuning. When the system degrades, nobody can say whether it is the retrieval, the rerank, the prompt, or the model.
Want this layer audited in your stack?
We run the K-Framework against your AI build and hand you the gap list, ranked by what it will cost you in production.