Why a ranked list fails
Ask a retrieval system how a memory note relates to your query and it hands back 0.82.
Two notes can earn the same score for opposite reasons: one is direct but obsolete; another is less direct but independently corroborated; two more can both be “relevant” and flatly contradict each other. A ranked list collapses which-is-which — the very thing that decides whether you can safely use the result.
And there’s a failure a per-document score can’t see at all: the pieces are each fine, but the answer they compose isn’t supported — the definition is there, the current status is there, yet nothing actually backs the causal link between them. That’s a shift from “is this document suitable?” to “does this evidence hold up when assembled?” — and it needs its own view.
A concrete walk-through
Query: “Does City X’s congestion charge cut traffic, and is it still in force?” Four notes come back — and a ranked list would show four almost-identical scores that hide everything that matters.
Claim coverage — no single note answers both halves
| Note | reduces traffic? | still in force? |
|---|---|---|
| A 2019 study | yes · stale | — |
| B dept page | — | yes · current |
| C op-ed | disputes | — |
| D explainer | — | — |
0.79 / 0.77 / 0.74 / 0.71 — placed so the differences are what you see first. B sits near center for the still-in-force half, yet the coverage grid shows it says nothing about the effect half. Proximity to center never means complete answer coverage.The thesis
Relevance is a reflection coefficient — complex, so it carries both how-matched and how-wrong; and reference-relative, so context reshapes the whole space rather than rescoring each item.
-
Reflection, not distance.
Standard retrieval treats a note as a point and asks “how far.” This treats it as a load seen from the query and asks how much of the query’s energy it absorbs versus bounces back — relevance as impedance match, reference-relative by construction. (maximum-power-transfer ≡ matched filter — real math)
-
Relevance is 2-D: magnitude + phase.
A single score is a projection that throws away the failure direction; a complex value keeps “how relevant” and “what kind of wrong” in one object. This half has production evidence — RotatE and ComplEx already use phase to carry relational meaning.
-
Context is a group action, not a recompute.
If context-conditioning is a Möbius transform, then “context A then B” is one composed transform, and it’s invertible — relevance-under-context has algebraic structure instead of a fresh score each turn.
Status — a thesis, not a product claim
- Representation
- Coherent and original.
- Ranking
- No — |Γ| provably collapses to cosine + recency. This is a reframing, not a “beats the baseline” claim.
- Composability
- Aspirational. An out-of-sample test showed the group-action leg holds only while context-conditioning stays close to a true Möbius map — a narrow band. So: one proven leg (phase), two elegant-but-unproven (reflection-framing, composability).
One proven leg, two beautiful hypotheses. That’s a theory — stated as one, not smuggled in as a result.
The move
Borrow the Smith chart from RF engineering. It maps a component’s impedance to a reflection coefficient Γ — loosely, how much of a signal bounces back instead of transferring — inside a unit disk: matched at the center (Γ=0), fully mismatched at the rim (|Γ|=1). The borrowing here: treat each note as a load seen from the query, feed a semantic-alignment score and a staleness score into an analogy-inspired mapping — alignment in the role of resistance, staleness in the role of reactance — form z = R + jX, and compute Γ = (z − 1)/(z + 1).
The honest limits, stated plainly: the embeddings are the retrieval engine. This geometry does not rank better than cosine, and it is not RF physics applied literally — cosine is not a resistance, staleness is not a reactance. They are display encodings. The entire value is in what the encoding lets a person see and do, not in any claim that a note has a true impedance.
The visual grammar
- Radius — a monotonic display transform of semantic alignment (not truth, provenance, or answerability); freshness is encoded separately, never silently mixed in.
- Angular sector — a declared mismatch category (“aging” vs “unresolved”), not a natural continuous dimension. It’s an ontology you choose, not one the geometry discovers.
- Beyond the rim — an authored alert state for detected contradiction, because “disagrees with the query” is not the same as “weakly relevant.” Not an automatic consequence of the transform.
- Locus under controlled paraphrases — re-ask in meaning-preserving ways; each note traces a path. A tight locus is a hypothesis that the note is stable across wordings; a wide one flags phrasing-dependence.
- Claim coverage — for compound questions, a note may support one component while leaving another unsupported. Proximity to center never means complete answer coverage — the state the companion reflectogram exists to make legible.
What you can do with it
Inspect — see, at a glance, why a candidate is where it is: on-topic-but-stale, relevant-but-open, contradictory, or generic. Revalidate — for a note drifting into the aging sector, re-check and re-date it and watch it move back toward center; different claims should age at different rates. Reconcile or bridge — a contradiction past the rim opens an evidence branch that ends honestly as revised, disputed (both kept, never averaged into a false middle), or unresolved — never quietly tuned to the center; and a set of individually-imperfect notes can be pulled inward by one shared context, shown with an audit of what that context costs.
Guardrails — a lens, not physics
- The positions are authored encodings, not discovered laws — every one is a design choice.
- The display must show its inputs, transforms, uncertainty, and provenance; a stability signal without provenance is an agreement signal, not an echo-chamber detector.
- RF theorems motivate questions; they do not validate retrieval claims. Keep the names as interface semantics — presented as mechanics, the whole thing collapses.
The transferable idea
Don’t represent retrieval as a single relevance number — represent it as an inspectable diagnosis of why a candidate succeeds, fails, or is unsafe to use.
This is the portable takeaway; the thesis above is the sharper, riskier version of the same claim — that the diagnosis has a natural shape, a complex value rather than a scalar. Semantic fit, temporal validity, agreement/conflict state, coverage of the answer’s claims, robustness to query variation — expose which dimension drove the decision and what intervention would improve it. That’s retrieval as state-estimation and diagnosis, not nearest-neighbor lookup — and it transfers to RAG evaluation, search debugging, knowledge-base maintenance, and agent observability.
How you’d know it helps
The test isn’t “is the metaphor elegant?” It’s “does it produce better decisions?” Build 10–30 deliberately constructed cases — current-and-supported, close-but-stale, two contradictory with unequal provenance, a paraphrase-brittle top hit, a multi-hop answer whose pieces are fine but whose composition fails, an echo chamber tracing to one unverified source. Show each three ways — plain ranked list, list-plus-metadata, and this visualizer — and measure whether people spot the failure mode, pick the safer evidence, and choose the right fix faster, while watching the opposite risk: over-trusting a display because it looks mathematical.
The Smith chart didn’t turn out to be a better ranker, and it was never going to be. It turned out to be a good excuse to stop reducing retrieval to one number — and to start treating “why did this match, and should I trust it?” as something you can see, interrogate, and repair.
↗ Open the interactive Γ-Locus Explorer
Appendix — further instruments (speculative)
These map to real RF theory but are, so far, thought experiments — kept out of the main argument on purpose: an L-network of “moves” (reaffirm / resolve / bridge) that tunes one note toward center; a Bode–Fano-style broad-vs-deep frontier (a note can be exquisitely matched to one phrasing or moderately matched to many, rarely both); multi-load “broadband” matching for whole clusters; and a time-domain reflectogram that marks where answer-building breaks. Each is an analogy for intuition, not a proof about retrieval, and each earns its place only once its inputs, transform, and the decision it improves are all spelled out.
A design exploration built as a personal thinking tool. Synthetic examples throughout. The geometry is a lens; embeddings do the retrieval.