An assistant finds a supplier's address in an old contract. It also finds a newer address in a verified account record. Both are relevant. Which one should it use?
A useful answer needs more than the closest text match. It needs to know where each address came from, when it was observed and what makes one source suitable for this decision.
That is where the database becomes part of the reasoning system. 8DB, our omnimodal database, includes structures for provenance, multiple confidence dimensions and explicit knowledge states. Those structures let an application retrieve the evidence around a fact and choose an appropriate next step.
For AI builders, the opportunity is practical: fewer assumptions about what a retrieved value means, and a clearer path from an answer back to its support.
Give retrieval an evidence packet
Retrieval-augmented generation connects a model to external information. The original RAG research demonstrated the value of combining learned models with retrieved material for knowledge-intensive tasks. The next application-level question is how to judge and use what was retrieved.
8DB's typed result structures can carry a value together with confidence, provenance and coverage information. Its confidence representation separates dimensions such as recency, provenance pedigree, agreement and empirical support. Keeping these dimensions available gives an application more to work with than a single relevance score.
Consider the address example. A support assistant may need the currently verified contact address. An investigator may need both addresses and their history. The right response depends on the question; silently selecting one value for every use would discard useful information.
The illustration below describes that decision. The records are hypothetical, and the routes shown are application choices rather than measured model behavior.
8DB · Architecture in practice
The same question can require three different actions
Illustrative question: Which address should we use for this supplier? No numerical confidence scores are implied.
- RecordWhat was found?
The candidate address and the source that supplied it.
- EvidenceWhat supports it?
Recency, provenance, relevant confidence dimensions and known coverage.
- Application policyWhat is sufficient?
Choose the next action for this task using the supplied evidence.
Use a currently verified address and retain its source.
Show the disagreement and identify what needs review.
Seek the missing evidence before drawing a conclusion.
Keep uncertainty useful
A confidence value needs a defined meaning. A score based on recency is not automatically a probability that a statement is true. Agreement can also reflect several records repeating the same upstream source.
Preserving the dimensions helps a team set rules suited to its task. It might accept a current authoritative address, qualify an answer based on older material, or request review when independent sources conflict. Where the product needs calibrated probabilities, the team must establish that calibration against an appropriate evaluation set.
The advantage is that uncertainty becomes something the application can act on. It can explain why an answer is qualified, identify the missing evidence and direct the next lookup.
Let missing information change the next step
"No value returned" is an incomplete instruction to a reasoning system.
8DB's typed views distinguish a present value, an expected absence with a reason, and an unknown state with coverage information. Those distinctions can support very different actions. A fact that does not apply needs no further lookup. A fact that is unknown may justify more research. A superseded value calls for its replacement.
The access layer must also decide what can be disclosed. An application should not reveal the existence of a protected record merely to explain why it could not return it.
For an agent, a well-scoped gap can be useful output. It tells the next step what is missing without turning incomplete retrieval into a confident claim that nothing exists.
Knowing what is missing can also help a team spend its effort wisely. An application can reuse an adequately supported result, seek a specific missing source, or leave the question unresolved. Measure the cost of those decisions across the complete workflow, alongside whether they were correct.
Make the answer reviewable
Provenance supplies the route back to the evidence. The W3C PROV standard provides an established model for describing how information was produced and related. 8DB's contribution here is to make provenance available alongside the result and other reasoning-relevant state in its database structures.
That can shorten an analyst's review: inspect the source, follow the transformation and see which conclusion depends on it. A source's identity and an intact lineage still require an assessment of whether the source is correct and applicable. Keeping that distinction visible makes the review stronger.
Our entity-resolution case study shows why retaining support matters when a conclusion is challenged. The same principle applies to a supplier address, an operational observation or a proposed answer in an agent workflow.
Test the decisions, not just the retrieval
An evaluation can begin with three versions of one question: a well-supported answer, conflicting evidence and incomplete coverage. Hold the question constant and change the evidence. Does the application answer, qualify or investigate for the right reason? Can a reviewer recover the source behind its choice?
8DB provides data structures that help build this behavior. The model, retrieval adapter and application policy must use them correctly; the presence of metadata alone does not establish a reduction in hallucinations.
The commercial value is a retrieval system whose outputs are easier to explain and act on. A team can see what supports an answer, where uncertainty remains and what evidence would help resolve it.
Bring us one question where a relevant answer is not enough. An example of conflicting or incomplete evidence is enough to define a useful first evaluation.
Evidence scope Implemented data types · Source review · Illustrative retrieval example. Explore the evidence.