Concept space
Every product and query is projected locally through a committed 24-concept lexicon. No endpoint or API key is read for this path.
01 · Space selection
Choose once, report it every time
EmbeddingSpace resolves one vector space for the
process. The default and automatic choice is the authored concept space. The real-vector
path is explicit because a run must not silently change meaning when credentials happen to
be present. It also requires --confirm-paid because every query may be embedded
by a live provider. Provisioning and the exact four-variable contract live in the
Microsoft Foundry live-run setup.
Real-vector space
The product index is committed, but a new query must be embedded into the same model space. A product identity probe checks that the live query embedder and index agree before use.
02 · Hybrid retrieval
Meaning and exact identity need different legs
Dense search helps when the customer's wording differs from the catalogue. Lexical search protects model numbers, GTINs, brands, and literal product facts. Their scores are not comparable, so VITRINE fuses ranks rather than raw values.
| Lexical signal | Current weight/boost | Purpose |
|---|---|---|
| Name | 3.0 | Product identity and strong title terms |
| Brand | 3.0 | Exact brand preference |
| Specification value | 1.5 | Mounts, capacities, dimensions, and compatibility values |
| Specification key | 1.0 | The kind of fact being requested |
| Description | 0.75 | Recall without letting prose dominate identity |
| Model-number match | +6.0 | Dense retrieval is weak at arbitrary identifiers |
| Exact GTIN | +12.0 | An exact barcode is an identity, not a semantic hint |
03 · Authored concept vectors
Deterministic meaning, with a visible ceiling
The key-free embedding source is a small authored projection—not a miniature claim of general language understanding. Text is tokenized, phrases of up to three tokens are looked up, concept weights accumulate, and the result is normalized.
What it is good at
Stable demonstrations, known retail concepts, understandable explanations, and repeatable offline verdicts. Artifact bytes differ because receipts deliberately contain run IDs and timestamps.
What it is bad at
Unanticipated vocabulary, linguistic nuance outside the lexicon, and any claim of broad semantic generalization.
Why zero is visible
An unrecognized query has no dense signal. The retriever records a degraded reason rather than ranking 99 identical zero cosines.
04 · Space-dependent operating points
A threshold belongs to a vector space
VITRINE carries separate rows because the numeric shape of an authored 24-dimensional space is not transferable to a 1,536-dimensional provider space. These are the values in current code, not claims of universal accuracy.
| Resolved space | Dense floor | Attribution floor | Primary confidence | Secondary confidence |
|---|---|---|---|---|
| Concept | 0.280 | 0.200 | 0.703 | 0.455 |
| Real vectors | 0.223 | 0.221 | 0.520 | 0.437 |
Do not compare scores across spaces as though only the vector length changed. Query embedding, cosine distributions, dense cut behavior, attribution, and confidence routing all change together. The runtime prints the resolved space so the provenance cannot disappear.
Calibration limitation: these are committed operating constants, not a reproducible public calibration package. The original derived calibration JSON and split are not present in this sanitized repository. The recorded held-out bands did not validate every row: the concept dense floor was cleared by roughly 57% of arbitrary products and the real-vector dense floor by roughly 24%. Treat these cuts as sample operating points that need re-derivation on governed production data, not as universal similarity thresholds.
05 · From hit to recommendation
Retrieval is necessary, not authoritative
A fused hit is only a candidate. Recommendation authority arrives later, after the subject can connect the product to a customer signal and to a resolvable catalogue fact.
- Pre-filter
Market, category, attribute, ownership, and exclusion constraints are applied before each leg ranks.
- Retain provenance
Each hit keeps dense rank, lexical rank, raw diagnostic scores, and the best matching catalogue line.
- Attribute to an interest
The subject must identify which independently derived or explicitly stated customer need the product answers.
- Resolve catalogue evidence
The cited review or attribute must resolve against the selected SKU. Plausible prose is insufficient.
- Run guardrails
Sensitive-category, confidence, candidate-containment, compatibility, ownership, and price/stock checks may drop or demote it.
- Re-read price and stock
Presentation uses catalogue authority at render time. A price stated in model-authored reason text is rejected.
06 · Interpretation limits
What this subsystem does—and does not—show
Offline reproducibility
The default search can run entirely in-process over committed product data and the authored concept projection. Its graph, counts, floors, and failure semantics are testable without credentials.
Stated constraints
A predecessor Galaxus-themed synthetic-sample run measured 0.889 live versus a tag-join oracle's 1.000 at matched k. It is not a Digitec Galaxus production or business metric, and this deep dive does not regenerate that paid measurement. Inspect the bounded receipt →
Next-purchase prediction
Four informative pairs make the minimum attainable two-sided p-value 0.125. Better retrieval mechanics do not repair an underpowered design.
Model value
A tag join reaches 1.000 on several questions with no model calls. Hybrid retrieval and a richer architecture must earn their complexity on matched comparisons.
Security boundary: the offline VITRINE
injection gate runs AgentEval's real RedTeamRunner with direct, indirect, and
tool-output attacks against deterministic guarded and deliberately vulnerable calibration
arms; its poison-dependent forbidden call is not Robin. The separate, explicitly paid
Eval 06 plan runs jailbreak and canary-backed system-prompt-extraction probes against fresh
real Robin targets and persists a redacted census. The offline gate is not a live-agent
resistance rate, and neither lane is a retrieval-quality rate. See the
evaluation protocol for their exact evidence and cost
boundaries.
For implementation-level details, start with
architecture reference and the source files under
src/AgentEval.VitrineDemo/Retrieval.