Ask in plain language.
Get governed evidence
ESRA — Enterprise Semantic Retrieval Agent — lets an AI client ask a question in plain language and get back governed evidence from your own enterprise databases, with no vector database, no embedding pipeline, and no copy of your data unless you ask for one. It runs inside the AI API Mapper platform and is included on every plan at no additional cost.
Retrieval that runs where your data already lives
ESRA is published to AI clients as an ordinary MCP tool through the ApiMapper Runtime, so everything that already governs a tool call — authentication, access policy, audit, rate limits, tenant isolation — governs a retrieval call unchanged.
Retrieval runs on the native full-text search engine of the database itself. Nothing leaves the source system to be indexed, ranking traces back to term matches in named fields rather than a distance in an opaque space, and record-level security is enforced by the database, inside the query.
How a retrieval call works
Seven stages, every one of them bounded and traceable. The AI model proposes possible search strategies; the platform's own deterministic logic decides what actually runs.
Resolve authorized sources
Authorization happens first, always. A source the caller may not reach never appears in a prompt, a log, a diagnostic, or a response.
Select the relevant ones
Mandatory sources are always searched. Optional ones are chosen by a service model from business descriptions only — never a schema, never a host, never a credential.
Plan several searches
Per source: an exact phrase, a broad keyword search, synonyms, acronym expansion, and a recency-weighted variant — so one phrasing of a question doesn't decide the answer.
Validate every plan
The model produces structured intent, never SQL. Each plan is validated deterministically before it can run, so a plan that fails validation never reaches a database.
Execute under bounds
Parameterized full-text queries run concurrently against a read-only database principal, with per-source timeouts, row caps, and circuit breakers.
Normalize and deduplicate
Results from every engine are merged into one shape and deduplicated, preserving which queries found each record.
Rerank and return
A semantic rerank produces the Top-K evidence. Where the exposure permits it, ESRA composes a grounded answer from that evidence and nothing else, citing the records it used.
No evidence means no answer rather than an invented one — an answer with nothing to ground it is exactly the failure mode grounded answers exist to prevent.
Key capabilities
Six areas that make plain-language retrieval something an enterprise can actually put in front of an AI client.
No vector database
No embedding job to run, no vector store to secure, and no staleness window between a record changing and an index catching up. Your data stays in the system that owns it.
Six engines, one contract
PostgreSQL, SQL Server, MySQL/MariaDB, Elasticsearch/OpenSearch, MongoDB, and Oracle Text — each through its own native full-text features. No term the model produced ever becomes query structure, only ever a value inside it.
The model proposes, the platform decides
Every model-assisted stage has a deterministic fallback: if the model is slow or unavailable, retrieval still works and the response is marked degraded rather than failing outright.
An exposure can only narrow
A source permits, a capability binds a subset, an exposure publishes a subset of that. Attempting to widen a limit is a publication error with a named reason — never a silent clamp.
Separation of duties
Holding the connection secret, modeling what is searchable, authoring the access policy, and publishing to an audience are four different rights. No single role can complete that chain.
Guided onboarding and Try a query
A wizard reads the database catalog and proposes what to search. Once published, Try a query runs the real pipeline and traces every stage, so "why did that return nothing" is answerable before a caller ever asks it.
What a caller never sees
Never in a response: SQL, connection details, credentials, stack traces, provider diagnostics, or any field not explicitly marked returnable. Record-level security is applied as a parameterized predicate inside the query, and a missing claim is fail-closed — the query is not executed and the source is dropped from the invocation. ESRA never substitutes a default and never omits the predicate.
Ready to let agents read your data safely?
Get a technical walkthrough with our team — or explore the ESRA documentation at your own pace.