No stale index
A retrieved fragment is read from the source that the application also serves.
AI Retrieval
Documents, embeddings and rows in one query.
Retrieval is the working part of most AI systems: find the right fragment, then answer from it.
A document that is found by similarity and read from the source that still exists.
Similarity search and context assembly over documents, records and events.
Retrieval quality depends on what is searchable and whether it is current. PLOMID treats similarity as an access path over the same data that serves the application, so a fragment found is a fragment that still exists in the source. Similarity is designed as an access path rather than a second copy of the data.
A document that is found by similarity and read from the source that still exists.
Environment
The documents
Manuals, tickets, contracts and attachments already in the layer.
JSON / documents · Objects
Two estates, drawn.
The same shapes, held two ways. The difference is not the storage, it is where the agreement between them lives.
Separate systems Copies kept in step
One layer One plan · one contract
What this workload moves.
Every shape below is a surface of the layer, not a format to be converted into one. They are read and written together.
- Vector Similarity and semantic retrieval
- JSON / documents Documents and nested objects
- SQL Records, keys and joins
- Objects Large assets with queryable metadata
How the workload reaches the data.
The models this workload names, the path a request takes through them, and the surfaces that speak to the layer.
AI Retrieval · architecture
Workload
The shape of the work itself.
- Source
- Represent
- Retrieve
- Assemble
Data models
The models this workload names.
- Vector
- JSON / documents
- SQL
- Objects
The layer
One path from a request to the data it names.
- Planning Predicates narrow the work before it runs
- Execution Rows, fields and windows answered together
- Transactions Readers and writers do not block each other
Surfaces
How the workload reaches the layer.
- SQL surface The query language the layer is documented in
- Applications Services and jobs reading and writing as they run
- Analytics & AI clients The same layer, the same access path
The work, stage by stage.
Choose a stage to read what happens there, or a shape to see every stage that handles it. Nothing on the map is a private interface.
PLOMID · AI Retrieval source · represent · retrieve · assemble Select a stage
Stage
Source
Documents, records and attachments in the layer
JSON / documents · Objects · SQL
01
Source
Documents, records and attachments in the layer
- JSON / documents
- Objects
- SQL
02
Represent
Embeddings held beside the source data
- Vector
03
Retrieve
Candidates found by similarity, filtered by structure
- Vector
- SQL
04
Assemble
The context returned with its sources
- JSON / documents
What changes when the data is in one place.
Stated as properties of the system rather than as results we cannot measure for you.
Filters plus similarity
Structured conditions narrow the candidate set before similarity ranks it.
Attachments are first class
Documents and their metadata are queryable rather than being an opaque store.
Provenance returned
Each result can be traced to the record or document it came from.
What this is built on.
Retrieval is only as useful as its provenance. These are the parts that carry it.
- Retrieval as an access path → Similarity search arrives as another way into the same data rather than a second copy to keep in step.
- Documents beside rows → Fields that keep changing shape stay queryable instead of being exported into a separate document store.
- Objects with queryable metadata → Large assets are found by a query against their metadata rather than by a naming convention.
- One data layer → Rows, documents and time-ordered events live in one system, so a question is asked once instead of once per store.
Workloads that sit beside this one.
Where this shows up.
Derived from the industry pages that reference this solution, so the two directions of the relationship always agree.
- Fintech → Ledgers, balances, product events and the relationships behind them.
- E-commerce → Orders, catalogues, inventory, sessions and fulfilment events.
- Healthcare → Clinical records, encounters, observations, orders and results.
- Life Sciences → Experiment records, instrument data, sample lineage and publications.
- Pharmaceuticals → Trial data, batch records, regulatory submissions and pharmacovigilance.
- Scientific Research → Instrument output, simulation results, reference material and datasets.