One Endpoint, Eight Answers

We built the search surface that made every agent a better searcher.

The problem was obvious

The platform could do everything: vector search, fuzzy text search, code graph exploration, memory recall, live sessions, recursive reasoning — eight modalities, each real, each living in its own world with its own API, its own auth, its own pagination, its own error shape.

Agents could only see what one modality offered. Ask "search the codebase for this bug pattern" and you got vectors — because vectors were the only thing one call away. The other seven answers existed, and were invisible from where the agent stood. That is the shape of a platform that is feature-complete and product-empty: everything works, and nothing is reachable.

So we built the missing surface.

The solution was simple

One endpoint that speaks all eight languages. aither-retrieval is a single service that turns the platform's existing substrate into one search API — one credential, one envelope, one pagination contract:

• Vector — the knowledge store, tenant-scoped, fail-closed
• Text — Postgres tsvector + trigram similarity, weighted and ranked
• Hybrid — all of it fused: normalized scores, weighted, deduplicated
• Graph — knowledge-graph neighbors and paths
• Code — the code graph: who calls this, what does this touch
• Memory — the memory store's recall
• Session — live REPL sessions, executed against the platform
• Recurse — top-k retrieval, then recursive reasoning over the results

Twenty-eight routes in all, including the pieces that matter most in a search API: opaque cursor pagination (keyset under Postgres, with the cursor bound to the tenant who minted it — hand it to a different tenant and it is rejected), idempotent writes with real conflict semantics (a replay returns idempotent_replay: true; a conflicting payload under the same key returns 409), a collection-points API for vector-shaped workloads, and a date-versioned contract — every response carries its API version and a request_id, and every error comes back as a structured RET-xxxx envelope with a reason and a message, never a bare 500.

It is public — retrieval.aitherium.com — authenticated by tenant bearer token, so a customer's agent can query the same surface from anywhere. And it ships with a skill that teaches a stranger exactly that: the base URL, the credential, the cursor contract, the idempotency rules, the error-code table.

Built to stay honest

A search surface is only as good as the contract it keeps, so the service treats its own contract as part of the product. Every response is envelope-shaped and versioned, every error is a code you can program against, and every store write is idempotent by construction — the same payload twice is a replay, not a duplicate, and a conflicting payload is a loud 409, never a silent overwrite. Tenant scoping is fail-closed: a request that names no tenant, or names one it cannot prove, gets nothing — never everything.

That contract is asserted continuously by gates that run unattended and page on violation, not on a schedule of good intentions:

• a contract gate — route order, error-code registry, API version and event protocol stay exactly as the surface declares them
• a live probe — an authenticated round-trip against the running service that asserts the envelope, the idempotent replay, and the conflict path
• an env-plane gate — every deployment unit's credentials file exists and carries the derived values its service declares

Two of the three run fleet-wide every four hours. A surface that isn't continuously verified is a surface that will quietly rot. Ours answers the same way tomorrow as it does today.

What it means

An agent that can ask one question and get eight answer shapes is a different kind of agent. It stops settling for the modality that happens to be wired and starts asking the question the way a human would: what do the code graph, the memory, the transcripts and the fuzzy text search each have to say about this?

The platform had all of it. It just had no surface. Now it does — one endpoint, eight ways to answer, public, versioned, and continuously verified. That is the whole story, and it is a good one.