Credits are a unit on someone else's meter. The post gives consumption, 30,000 to 40,000 per simulation, and no rate [2][25], so the figure that transfers is the multiplier rather than the money: one run a day for a year is 10.95 to 14.6 million credits [23]. Against a weekly scenario that estimate is noise. Against a daily batch with reads served on top of it, the provider's availability, billing model and isolation semantics are now inside your read path [4].
What made the swap cheap was the read profile, not the database. The framing invites a bad design: reproduce the provider's graph product inside the gateway, then keep a second graph engine running beside MiroFish [11]. Two graph engines in one service is one more than a launch scope can staff. What actually shipped was three tables, graph_nodes for entities mirrored from the upstream graph, graph_edges for typed relationships, agent_episodes for round actions and their content vectors [15], sitting in the same transactional store as the existing records [9]. pgvector answered the semantic query without adding a network hop [10].
The isolation work is where owning the store pays. Both graph tables carry tenant_id and simulation_id, and unique constraints stop the same upstream entity or episode being written twice for one simulation [13]. Two tenants can hold a node with the same upstream identifier without either seeing the other's row [14]. Behind a service boundary that behaviour is whatever the provider implements, and you find out which it is during an incident, not during design [4].
The embedding path is built to make one coupling loud. Width is 384 because the in-process Xenova/all-MiniLM-L6-v2 emits exactly that shape, with mean pooling and normalized output [16]. The table definition, the embedding constant and the tests all name 384 [18], which is the author's argument that a silent dimensional mismatch is worse than a migration you planned [18]. The extractor initializes lazily behind one shared promise, so requests that never touch semantic search do not load the ONNX model and two concurrent callers do not start two loads [19]. The cost is cold-start latency on the first embedding request, accepted because this gateway runs long simulations, with pre-warming at startup held in reserve if search ever becomes an interactive first screen [20].
Two conditions carry the whole build case: the required traversals were bounded, and the semantic search was local [21]. Check both against your own workload before the credit estimate means anything, because the estimate describes the shape of this author's simulation, not yours. If your reads walk unbounded paths, or your embedding model is chosen by a vendor who updates it, the arithmetic above stops being the relevant arithmetic.
