Use Case — RAG

Find where your RAG pipeline breaks — retrieval, reranking, or synthesis.

Hallucinations don't appear from nowhere. They come from bad retrievals, ignored context, or prompts that drift. Zespan shows you exactly which step is failing.

Start free →For Teams building retrieval-augmented generation systems

The problem

01

Hallucinations with no trail

You can see a bad output but not the retrieved context that caused it. Was it a poor retrieval, a bad prompt, or the model ignoring context? You can't tell.

02

Retrieval quality is a black box

You don't know your average retrieval score, how many chunks are retrieved, or whether reranking actually improves results at scale.

03

Quality degrades silently

Faithfulness scores slip as your document base grows or prompts change. You find out from user reports, not metrics.

How Zespan helps

1

A RAG pipeline's steps are separated, not collapsed into one call

Query encoding, vector retrieval, reranking, and synthesis each trace as their own timed step. A hallucination stops being a single opaque LLM response and becomes a pipeline you can actually inspect stage by stage.

zespan.com
Zespan trace explorer filtered to RAG operations showing latency, cost, and status
2

You can see whether retrieval or synthesis is where a response went wrong

The waterfall shows retrieval and synthesis as separate spans, each with its own timing, and retrieval spans carry the chunk scores that entered the context window. A low-score chunk that made it into a final answer is visible as the specific cause, not a guess.

zespan.com
Zespan span waterfall showing RAG pipeline with retrieval, reranking, and synthesis spans
3

Every response is scored for whether it's actually grounded in what was retrieved

A faithfulness evaluator runs automatically on every response: a score near 1.0 means the answer is grounded in retrieved context, near 0.0 means the model answered from memory instead. You find the hallucinations by the score, not by a user telling you the answer was wrong.

zespan.com
Zespan evaluations showing faithfulness and relevance scores trending over time
4

A low score comes with the evidence, not just the number

Behind any low-faithfulness trace is the actual question, the chunks that were retrieved, the answer that was generated, and the reasoning for the score. The gap between what was retrieved and what was said is what tells you whether the retrieval was bad or the model ignored good context.

zespan.com
Zespan evaluation detail view showing faithfulness score with retrieved context and judge reasoning
5

A hallucination you found once doesn't get to happen the same way twice

A failing trace becomes a dataset entry built from the real failure, not a hypothetical test case. The next retrieval config or prompt change runs against that dataset before it ships, so a fix that quietly reintroduces the same hallucination is caught before a user sees it again.

zespan.com
Zespan trace-to-dataset showing a failing RAG trace added as a regression test case

Zespan features used

Tracing & ObservabilityDeep dive →EvaluationsDeep dive →SimulationsDeep dive →Prompt ManagementDeep dive →

Start free — 50K traces/month, no card needed

See every agent decision, tool call, and handoff in production. Setup takes a few minutes.

Start free →
← Back to home