Status: accepted · Date: 2026-09-10

#572 is lot 2 of #443's Phase 5, and #443 calls it the strategic argument of the whole roadmap: a hybrid vector store with no external database, where every other provider in the Semantic Kernel ecosystem needs a service running somewhere. Its acceptance criteria ask for this record before code, on two questions — where the conforming types live, and what the new inter-package edges cost.

Both are settled here. Nothing about the adapter's own shape is: UpsertAsync against an append-only EmbeddingIndex, and what a compiled Expression<Func<TRecord, bool>> filter means in process, are implementation decisions and belong to the code that makes them.

ADR 0074 requires a package's exported surface through a MetadataLoadContext, not its README. Microsoft.Extensions.VectorData.Abstractions 10.9.0, MIT, read on 2026-09-10:

exported public types	44
target frameworks	net10.0, net8.0, net462, netstandard2.0 — both of this repository's
package dependencies	exactly one: Microsoft.Extensions.AI.Abstractions 10.9.0
VectorStore	6 abstract members
VectorStoreCollection<TKey, TRecord>	11 abstract members, Name included
IKeywordHybridSearchable<TRecord>	2, one of which is the GetService the collection also carries

Every type #572 names is there, IKeywordHybridSearchable<TRecord> and HybridSearchOptions<T> among them. The single dependency is the version Lodestar.Extensions.AI already pins, so no version range has to be reconciled.

A conforming package carries an external dependency, so it is satellite tier rather than core (decision 0076). The name follows Lodestar.Extensions.AI and Lodestar.Extensions.MathNet: the interop tier is named for the thing it interoperates with, and it is the tier decision 0089 opened for exactly this — a dependency a core package refused.

The option that had to be refused is folding it into Lodestar.Extensions.AI, which is tempting for a real reason — that package already carries Microsoft.Extensions.AI.Abstractions 10.9.0, the one dependency this needs, so folding would add no new external dependency at all.

It is refused on a measured cost. Lodestar.Extensions.AI carries an edge to Lodestar.Onnx, whose whole reason to exist is Microsoft.ML.OnnxRuntime — 1.28.0 is a 132.7 MB package (139 145 017 bytes), almost all of it native assets for platforms a given consumer does not run on. Folding would put that on the restore path of every caller who wanted a vector store and no model, which is exactly the caller this lot is for. A satellite that shares a dependency with another satellite is cheap; one that inherits the other's dependencies is not.

Lodestar.Extensions.VectorData needs two new inter-package edges, each with a published floor in src/Directory.Packages.props, an EXPECTED entry in tools/check_nuspec_dependencies.py and an edge in tools/check_version_floor.py:

• → Lodestar.Embeddings, for EmbeddingIndex — the vector half. Available: 0.6.0 is published and exports it.
• → Lodestar.Text, for Bm25Index and RankFusion — the keyword half. Not available.

src/ reaches a sibling through a PackageReference on a published floor (decision 0012), and the same MetadataLoadContext reading applied to the published Lodestar.Text 0.5.0 finds 79 exported types and not one of them Bm25Index, Bm25Options, Bm25Idf, SearchHit or RankFusion. The tag Lodestar.Text/v0.5.0 was cut on #558, two days before #573 added them.

So this lot lands after a Lodestar.Text release that carries the keyword half, and not before. That is the same ordering #569 met and is not a new rule; it is recorded here because #572's own acceptance criteria list the edge without listing the release it waits on, and because the alternative routes are both worse:

• Reimplementing BM25 inside the satellite duplicates a published algorithm across two packages, which the repository resolves by extraction rather than by copying.
• Taking the keyword half as a caller-supplied delegate removes the edge and, with it, the claim: a store that only ranks by keyword if the caller wires the ranking is not the batteries-included hybrid #443 refuses to ship without. That is option B under another name, and #443 refuses option B outright.

Issue #572 records that "the API is asynchronous and IAsyncEnumerable-centred, which nothing in this repository uses today." Half of that is stale and the half that stands is the expensive half. async Task is already here — EmbeddingIndex.SaveAsync, EmbeddingIndex.LoadAsync and four persistence loaders — so the asynchronous entry points are ordinary. IAsyncEnumerable<T> genuinely is absent, and every Get/Search on this surface returns one.

It costs no new declared dependency: on netstandard2.0 the type comes from Microsoft.Bcl.AsyncInterfaces, which is already in the restore graph transitively, and Microsoft.Extensions.AI.Abstractions brings it on that target in any case. What it costs is a first idiom — how a synchronous in-memory index is yielded asynchronously without pretending to be asynchronous — and the code that introduces it owns that.

• A fifteenth package, third of the interop tier, and the repository's sixth and seventh inter-package edges.
• Lodestar.Text must be released with Lodestar.Text.Search before this branch can be green. Releasing it is a separate act with its own tag, and this record is what says why it comes first.
• The three gates apply unchanged: a sample per public type, a reference page per public member, an equivalence row. The equivalence row says there is no Python reference for an interface adapter and that conformance is proven by driving the abstractions as a consumer would, so the suite does not read as skipped.