Give agents evidence beyond search results

A list of relevant records is useful evidence for a focused question. A different question may ask about the structure of a collection: which concerns recur, how records group together, or what examples characterize each group.

Those questions need an appropriate view of the data. EigenLens begins with clustering and topic modeling, giving developers and agents tools for producing that view from existing embeddings. The resulting evidence can become an input to reasoning, generation, or an application workflow.

From an operation to an explanation

Consider an agent asked to explain recurring concerns in a defined selection of customer messages. It needs to understand the selection, run suitable analysis, and inspect the result before writing an explanation.

The current MCP tools support those building blocks. retrieve_records or register_records establishes a dataset snapshot. run_analysis accepts clustering and topic operations. get_job reports whether the work has completed. After success, get_summary, get_examples, and get_result_page provide ways to inspect the saved results.

The explanation can draw on several kinds of evidence together:

Keep these elements connected. A topic phrase needs source context, and an example needs selection scope to support a useful explanation.

Scope travels with the result

Suppose an application retrieves a bounded set of messages similar to a refund question, then clusters those messages. The analysis describes that retrieved neighborhood. Its groups may help explain differences within refund-related results, but their sizes do not establish how common those concerns are across all support traffic.

A complete collection selection supports a different interpretation, provided it matches the question. Preserve the distinction in ordinary language: “Within the retrieved records…” or “Across the analyzed collection…” A sample remains a sample.

Counts need the same care. Ten vector records might represent ten documents, several chunks of one document, or something else. The source IDs and parent references determine what the application can reliably count. An agent should carry that meaning forward instead of silently turning vector counts into customer or ticket counts.

Make interpretation inspectable

Clustering discovers geometric groups under a particular method and configuration. Topic extraction produces weighted terms from source text and those group assignments. A human-readable name for a group is a further interpretation.

An agent might suggest “account recovery” after reading terms and representative messages. The useful explanation includes why that label fits and points back to the relevant records. If examples conflict or the group is broad, the language should reflect that uncertainty.

Include unassigned records and diagnostics when they affect the conclusion. They may reflect diverse material, unsuitable embeddings, or limits of the method. Display coordinates can help visualize a result, but their projection is not proof of a business conclusion.

A broader evidence toolkit

Our longer-term vision includes more evidence views. Classification could provide assignments to known categories with supporting references. Anomaly detection could identify unusual candidates relative to an explicit reference population. Recommendations and diversity tools could help assemble useful candidate sets.

These are future capabilities, without dates or a fixed release order. They need suitable inputs, such as reference categories, an anomaly baseline, or eligibility rules. Today, developers and agents choose the selection and operations; EigenLens does not provide an automatic intent planner.

The broader direction includes reusable models, new-arrival processing, stable cases, budgeted queues, and feedback. The homepage examples simulate how an agent could use today's evidence: combine support volumes into a work plan, turn missing case records into a review checklist, or connect launch decisions with customer research. These are authored responses grounded in sample excerpts and counts. The separate BANKING77 comparison uses saved analysis. These examples do not establish evaluated domain integrations.

Judge the downstream result on the task

Additional evidence can help or distract. Hold the question and selection clear, inspect what each approach receives, and assess whether the final explanation is correct, useful, and appropriately scoped. Better answers are a hypothesis to evaluate, not a guarantee.

The first-analysis walkthrough shows the current calls, and getting started covers setup. Start with a workflow you can inspect from source records through the final explanation.