Simple Collections
Group governed Sources for ordinary enterprise search and retrieval while preserving identity, provenance, permissions, custody, freshness, and use boundaries.
Enterprise information, governed before it becomes an answer.
Indwel Collections preserve more than searchable content. They preserve how a domain is organized, how its units are addressed, what retrieval owes the Work, and what material may become Evidence.
Start simple. Add structure only where the domain has structure to preserve.
Collections share one architecture from ordinary enterprise knowledge through exact, contract-governed information domains. The difference is declared structure—not a different product.
Structured Collections
Declare the domain's meaningful units—sections, clauses, records, metrics, periods, events, transactions, symbols, operational windows, or other addressable families.
Governed Collections
Add contracts that define membership, semantic units, selector families, retrieval behavior, Evidence requirements, completeness rules, and proof obligations.
Exact Collections
When the domain can identify what a request owes, deterministic selectors resolve the exact expected units and the system proves whether they were fulfilled.
Search when relevance is the question. Select when the domain already knows the unit.
Indwel keeps probabilistic retrieval and deterministic addressability in the same information architecture without pretending they are the same operation.
Relevance retrieval
Lexical/full-text and semantic/vector retrieval find candidate context according to relevance. Hybrid retrieval combines those paths as part of the Collection's declared retrieval behavior.
Deterministic selection
Semantic selectors resolve exact declared units from known domain structure. A question about a clause, period, transaction family, record, or operational window can be mapped to what must be returned rather than merely what appears similar.
Exact fulfillment is provable.
The Collection contract can derive the semantic units a request requires, project the units actually returned, and compare the two before synthesis or action proceeds.
- Resolve the request.
Map the request to a selector and the declared semantic units the domain says are owed.
- Enumerate expected units.
Establish the exact ordered set, unordered set, range, child set, or bounded fulfillment requirement.
- Retrieve or select.
Run relevance retrieval, deterministic selection, or both according to the Collection contract.
- Project observed units.
Read the semantic identities actually present in the returned material rather than assuming successful retrieval implies completeness.
- Prove fulfillment.
Compare expected and observed units. Missing material is repaired, narrowed, or refused rather than silently treated as complete.
Retrieval does not make Evidence.
The information architecture keeps acquisition, custody, organization, retrieval, and Evidence admission distinct all the way into Work.
Enterprise systems
Systems of record, repositories, applications, APIs, files, databases, and web sources.
Sources
Governed information resources with identity, provenance, permissions, custody, freshness, and use boundaries.
Collections
Governed groupings and publications with structure, contracts, semantic units, and selectors.
Retrieval
Relevance search or deterministic selection produces candidate material.
Evidence
A separate governed admission decision determines what the Work may rely on.
That distinction changes what retrieval means.
RAG usually ends the information problem once useful context reaches the model. Indwel carries the information contract farther: the system knows what material is, what structure it belongs to, what was requested, what was returned, and what was actually admitted as Evidence.
Collections are part of the Indwel information architecture.
They sit between governed Sources and cognition so enterprise structure does not disappear at the moment intelligence begins.
Bring a real information domain.
Ask Indwel how a corpus, database, policy set, operational dataset, or governed publication should be structured before the Work relies on it.