Introducing EigenLens

If you already store embeddings, you have the starting point for more than a search feature. Those representations may also help you understand which records belong together, what themes recur, and how to approach a collection before you know exactly what to ask.

EigenLens brings tools for that work to your existing vector database. Our starting point is clustering and topic modeling: group compatible embeddings, inspect the language that characterizes those groups, and carry the evidence into an application or agent workflow.

Built by Eigenlake, EigenLens gives agents more than search: useful evidence today, and a reusable decision layer for the work ahead.

Your vector store remains the source system. EigenLens reads a selected collection or set of records, materializes a local snapshot, and runs analysis in its own deployment. Developers can use the Python interface directly; agents can work through the MCP server.

Begin with a question you recognize

Start with a question from ordinary work:

• Support and feedback: What concerns keep appearing in these customer messages?
• Knowledge work: What are the main groups in this document collection?
• Content: What themes run through this article archive?
• Commerce: What recurring experiences do buyers describe in these reviews?

Each starts with a defined selection of records. For feedback, the result might include group sizes, characteristic phrases, and representative messages. An agent can use that evidence to explain recurring patterns while linking back to the source records. These are example uses, not completed customer deployments.

What you can use today

The current analysis tools are clustering and topic modeling. Clustering groups compatible vectors using a bounded parameter search. Topic extraction uses the discovered assignments and source text to produce weighted terms and phrases. Results include source IDs, representative examples, counts, and information about the analyzed selection.

The distinction between terms and interpretation matters. A list containing “refund,” “pending,” and “days” can help someone investigate a group. Calling it “refund delays” is an interpretation to check against the actual messages. EigenLens gives you the material for that inspection.

There are adapters for Qdrant, pgvector, Pinecone, Weaviate, and Milvus. Their connection requirements and retrieval behavior vary, so the setup guidance should be your reference for a particular store. The self-hosted Docker package provides the service runtime; Python is another way to use the analysis building blocks.

For a practical starting point, follow the first-analysis walkthrough and the maintained getting-started guide.

The toolkit we want to build

Clustering and topics are the beginning of a broader direction. Classification could help assign known categories. Anomaly detection could surface records that deserve closer inspection. Recommendations and diversity tools could help select useful, varied candidates for an application.

Those are future capabilities, without dates or a fixed release order. Their usefulness depends on the task, embedding model, and supporting inputs: reliable labels for classification, eligibility rules for recommendations, or a meaningful comparison population for unusual records. Evidence beyond search explains how those outputs could support an application.

From evidence to recurring decisions

The longer-term ambition goes beyond a one-off analysis. We want EigenLens to maintain reusable models, apply them to incoming data, prepare coherent work queues and experiment batches, and update them with feedback. Agents and domain applications could consume those prepared results while keeping appropriate human review and action boundaries.

The examples on our homepage make the roles concrete. Banking groups help an agent propose a support work plan. Legal case summaries expose missing evidence and dates to verify. Launch slides and customer interviews combine into a handoff brief. EigenLens supplies groups, topic terms and source evidence; the agent connects those facts to the task and proposes useful next steps with citations.

The surrounding application prepares embeddings and source links. EigenLens groups selected records and extracts topic phrases; those results return to the agent through MCP. The agent interprets the evidence and prepares the final presentation. The three use cases use authored examples and illustrative counts; the separate BANKING77 comparison uses saved analysis. These frontend previews do not run models or demonstrate validated domain integrations. Persistent models, classification, diversity, and recurring jobs remain future direction. Today’s starting point is clustering and topic discovery.

Docker first, open source for the long run

EigenLens is being built as a permanently open-source project, available for anyone to use for any purpose. The Docker image is the first available distribution. Public source release is planned; the source repository is not public yet.

Our priority is helping people get useful results and learning where the tools fall short. If you have embeddings and a question about the patterns within them, start with your own data.