Govern Information by Meaning
A unified metadata foundation for SQL and unstructured enterprise information
AI systems do not experience enterprise information as a neat collection of databases, files and applications. These systems receive context assembled from many types of information like a table definition, a policy document, a paragraph from a contract, a message from a collaboration platforms or a chunk retrieved from a vector index.
Governance must accompany that information throughout its entire process. A control framework solely centred on files, tables, rows and columns loses precision when content is parsed, chunked, embedded and combined into model context. The enterprise needs a common way to describe what the information means, where it came from, how it is related and which policy applies.
That is the purpose of unified metadata.
Storage format is an implementation detail for AI
Traditional data governance grew around the systems in which information was stored. Database teams governed schemas, tables and columns. Content teams governed folders, files and document permissions. Search systems created another set of indices and access rules.
Generative AI crosses these boundaries in one request. A user may ask for the obligations in a supplier agreement and the payment history of the same supplier. The answer could depend on a document clause, a database field and a business definition maintainedelsewhere.
The model does not need those sources to share a storage format. It needs accurate, authorised context. The control layer therefore needs to understand the business meaning of each information object and preserve the relationship between the original source and the representation used by the AI application.
One schema, several types of information object
A unified metadata model can give every governed object a common foundation while retaining the details that make each object useful.
The common foundation should describe:
- a stable identity for the object;
- the information on source and asset location;
- the object type and its relationship to a parent asset;
- technical and semantic descriptions;
- sensitivity and business classifications;
- lineage and provenance;
- stewardship and lifecycle state; and
- the policy references that govern access or use.
Typed details can then describe the object more precisely. A SQL table has columns, keys and relationships. A document has a title, content type and source location. A chunk has a position within a document, extracted text and the metadata inherited or derived from its parent.
This structure avoids forcing every object into an identical shape and creates a consistent governance surface across different shapes.
SQL meaning should survive beyond the database catalogue
Technical metadata can identify a database, schema, table, column and data type while the AI applications also need the meaning behind those structures.
A field called `cust_seg_cd` may be technically clear to the system that created it and meaningless to everyone else. Semantic metadata can record that it represents a customer segment, identify the business definition, connect it to a steward and describe its relationship to other fields. Lineage can show where it came from and which downstream assets depend on it.
Kynexa currently provides SQL connectivity, metadata generation, semantic modelling, lineage and stewardship workflows. Its semantic modelling combines language models with human input to describe tables, roles and columns. SQL information handling remainsdistinct from vector retrieval. The semantic modelling of the SQL data helps language models resolve ambiguity and provide critical inputs to the system admin to set data access policy for different users in the enterprise.
That boundary matters. Metadata unification does not require every information type to use the same retrieval mechanism.
Unstructured content needs governance below the file level
A file is often too large to be useful as model context. RAG systems divide documents into chunks, create embeddings and retrieve the passages that appear relevant to a request.
The chunk becomes the unit that can influence an answer. It therefore needs its own identity and metadata while remaining linked to the source file. A governed chunk should retain enough information to establish:
- which file and source system it came from;
- where it appears within that file;
- which classifications and access rules apply;
- which semantic labels were extracted; and
- how it contributed to a retrieved result.
Kynexa processes unstructured information into chunks and embeddings, attaches metadata at chunk level and aggregates metadata at file level. Produktiv uses its own chunking strategy which preserves the structure of the document along with multimodal content, tables, equations etc.
Meaning provides the bridge to policy
Storage permissions remain important, but they rarely express every rule an enterprise needs for AI. Two documents in the same repository may contain different levels of sensitive information. Two columns in the same table may have different business meanings. A single file may contain one passage suitable for broad use and another that requires restricted access.
Semantic classifications give policy a more precise target. Administrators can define access according to a user's role and label information with sensitivity or user defined semantic classifications. The policy layer can then evaluate the information that an AI workflow is preparing to use.
This changes the unit of governance. The enterprise can apply a policy to the meaning and sensitivity of an information object while retaining its connection to the source system.
A practical metadata lifecycle
The useful lifecycle is continuous:
1. Connect to the approved source and identify the asset.
2. Extract technical metadata and preserve source identity.
3. Parse unstructured content or inspect structured definitions.
4. Create semantic descriptions and classifications.
5. Link files, elements and chunks to their parents and lineage.
6. Attach or resolve the applicable policy.
7. Use the governed metadata during discovery and retrieval.
8. Record changes, decisions and stewardship input.
Each stage adds information without breaking the chain back to the source. Human stewardship remains part of the system because automatically generated meaning can be incomplete or wrong.
What this foundation enables
A unified metadata layer gives several AI applications a common view of enterprise information. A knowledge assistant, discovery application and future agent workflow can refer to the same identities, classifications, lineage and policy references instead of rebuilding those controls independently.
Kynexa supports SQL and unstructured information through a common metadata schema. It also extracts syntactic and semantic metadata and supports custom metadata extraction. This creates the foundation for governed context and retrieval while keeping the boundary clear for governed unstructured RAG and governed SQL retrieval and discovery.
The architectural principle is simple. Govern the information that AI can understand and use, while preserving the source, meaning and authority behind it.