Domain Language

Index governs vector retrieval. Its language distinguishes stored corpus state, the contract under which a query may execute, the ranked result, and the evidence needed to explain or replay that execution.

flowchart LR
    corpus["identified corpus and vectors"]
    intent["execution intent"]
    contract["mode, contract, and budget"]
    plan["resolved execution plan"]
    result["ranked result"]
    evidence["artifact, fingerprints, trace"]
    replay["replay or comparison"]

    corpus --> intent --> contract --> plan --> result --> evidence --> replay

Backend selection is deliberately downstream of intent and contract. A store name cannot substitute for an exactness promise, loss profile, resource budget, or replay posture.

Corpus and Retrieval

Term	Meaning
document	source identity and text admitted to the corpus
chunk	document-owned text span used as the unit of retrieval
vector	dimension-checked numeric values owned by one chunk
model specification	model identity, dimension, vendor, and version
vector store	adapter that persists or searches vectors
exact retrieval	exhaustive scoring under deterministic execution
approximate retrieval	ANN candidate selection with declared quality and randomness bounds
result	ranked document/chunk/vector row with score and artifact identity

A vector store is an implementation choice. It does not define the execution contract, and “stored successfully” does not imply that exact or approximate query requirements can be satisfied.

Execution Authority

Term	Meaning
execution artifact	corpus and index definition bound to an execution contract
execution contract	deterministic or non_deterministic guarantee
execution intent	reason the caller requires exactness, reproducibility, exploration, or another posture
execution mode	refusal posture: strict, bounded, or exploratory
execution budget	declared limits on latency, memory, error, vectors, distance work, and probes
execution plan	validated algorithm and backend decision for one request
refusal	governed rejection because the declared contract cannot be honored

Intent explains why a loss profile is acceptable. Mode defines how strictly the runtime must enforce it. Contract states the promised behavior. These are separate fields and should never be collapsed into a backend name such as “HNSW mode.”

Approximation and Quality

Term	Meaning
ANN	approximate nearest-neighbor retrieval
candidate pool	vectors selected for later exact scoring or reranking
witness	exact comparison used to measure an approximate decision
target recall	requested quality threshold for ANN behavior
decision trace	recorded non-deterministic choices made during execution
low-signal refusal	rejection when score or distance evidence is too weak
replay-strict	policy that refuses replay after index or parameter drift

A witness measures approximation; it does not convert an approximate run into an exact run. A seed can make some random choices reproducible, but it cannot erase backend, index, corpus, or parameter drift.

Identity and Evidence

Term	Meaning
artifact ID	caller-facing identity of materialized execution state
correlation ID	interface lifecycle identity and on-disk run prefix
execution ID	runtime identity of the resolved plan and result set
fingerprint	content-derived identity for vectors, configuration, backend, determinism, or results
explanation	join from one result to document, chunk, vector, artifact, score, and execution
replay	re-execution evaluated under the retained contract and equivalence policy
comparison	structural and ranking differences between executions or artifacts

“Same results” means only the compared result projection matched. “Replayable” requires the retained artifact, contract, inputs, backend conditions, and fingerprints needed by the replay policy.