Horizon-RAGLook inside the answer.
Build retrieval around your own sources, then inspect how the evidence was selected. Turn retrieval from a hidden step into an experiment you can understand.
AI enabled · Supported by your background knowledge
Coming soon, watch this space.
Coming soon, watch this space.
What belongs
in the answer?
Give more weight to meaning or to exact words. Watch the evidence order change.
A local illustration of weighted ranking, using fictional passages.
Practical Wisdom, in practice
What makes Horizon-RAG different.
Feature 1 of 6
A mixing desk for retrieval
Make the trade-offs explicit.
Adjust six weights: semantic search, keyword search, metadata, recency, authority and diversity. Set Top K, candidate pool and similarity threshold, and choose whether to transform the query or rerank candidates.
These controls let you ask what kind of relevance matters for the task. The similarity threshold applies to embeddings, not the final combined score—an important distinction when you interpret the results.
Compare a semantic-heavy search with an exact-term search over your own technical material.
View full size ↗Feature 2 of 6
Understand why a passage ranked where it did
Inspect the reason, not just the score.
For a selected passage, inspect base score, diversity penalty, final score and the contribution of weighted features. Ranked chunks show source, location, excerpt and context status.
Remove a passage from context while leaving it ranked, then rerun with the changed evidence selection. Restore it when you want to compare the effect of its inclusion.
Test whether a highly ranked passage is useful evidence or merely a strong keyword match.
View full size ↗Feature 3 of 6
A visible, versioned pipeline
Make configuration something you can revisit.
Build shows eleven stages from documents to answer. Inspect stage settings and export the pipeline definition. Change chunk size, overlap, separator and strategy, then re-index stored sources when appropriate.
Pipeline variants freeze the relevant settings as versions, and runs record the version used. Importing a pipeline definition creates a new version without replacing source material or past runs.
Keep a baseline pipeline, create a variant with different chunking and compare retrieval on the same question.
View full size ↗Feature 4 of 6
Compare runs without rewriting history
Preserve what happened before you tune the next thing.
Each run freezes its configuration and ranking. Select two runs to compare settings, top-ranked evidence, answers and recorded metrics. Saved projects retain runs, desk settings and presets.
A project file does not itself contain source texts or embeddings; index export and import handle the knowledge base separately. That distinction matters when moving the work to another computer.
Explain why a new configuration surfaced a different passage instead of relying on memory of the previous answer.
View full size ↗Feature 5 of 6
Inspect answers as claims with evidence
Keep support and uncertainty visible.
With an AI model configured, answers are expressed as short claims citing retrieved passages. A judge checks claim support, and quotations are checked against the passage text.
Explain shows support, flags, retrieved-but-unused material and skipped or failed steps. Without an AI model, the workbench still retrieves and ranks; it does not imply a generated answer exists.
Review the passage behind one claim and see whether it supports the wording or leaves a qualification unresolved.
View full size ↗Feature 6 of 6
Bring your sources—and keep them current
Retrieval should work with your material.
Add PDFs, Word files, HTML, Markdown, text, folders, web pages or knowledge packs. Sync updates tracked folders and pages, while re-indexing rebuilds embeddings from stored documents.
Web checks can be manual, daily or weekly while the app is open. The current preview does not claim a finished evaluation suite or Knowledge Galaxy; those remain later roadmap work.
Maintain a retrieval project around a changing folder of working documents and a selected set of official web sources.
View full size ↗Where your knowledge meets your judgement.
Broaden your Horizon.
Explore how Horizon-RAG can bring your proprietary knowledge into productive work.
Coming soon, watch this space.
AI and connected services may require your own provider accounts and separate usage charges. Feature availability depends on the product and build.
