​

Logs Explorer

Prerequisites: You have data flowing through Seer (see Quickstart) and understand Metrics.

​

Why Use Logs?

​

Two View Modes

​

Traces View (Default)

• Groups related spans by trace_id
• Shows the complete journey from query → retrieval → answer
• Displays final context metrics (what actually went to the LLM)
• Expandable tree visualization of span hierarchy

​

Records View

• Shows each evaluation as a single row
• Quick access to all metrics (recall, precision, F1, latency)
• Click any row to open the detail drawer

​

Filtering & Search

​

Traces View Deep Dive

​

Trace Table Columns

​

Metric Mode Toggle

• Final: Recall of the final context that was passed to the LLM
• Avg: Average recall across all spans in the trace

​

Expanding a Trace

• Tree lines show parent-child relationships
• Span badges indicate type (retrieval, rerank, llm_call, etc.)
• Final context badge marks which span’s context was used for the answer
• Delta indicators show how each hop changed overall recall

​

Trace Detail Sidebar

​

Records View Deep Dive

​

Records Table Columns

​

Sorting

• First click: ascending
• Second click: descending
• Third click: return to default (newest first)

​

Record Detail Drawer

​

Sections

​

Passage Visualization

• Position and ID (e.g., #1 • ID: doc-123)
• Gold badge if this doc is in ground truth
• Relevant badge if Seer marked it as supporting requirements
• Score from your retriever
• Full text of the passage
• Source if provided

• Green border: Relevant passage (correctly retrieved)
• Yellow border: Gold doc not marked relevant (missed)
• Gray border: Not relevant, not gold

​

Ground Truth Validation

​

Multi-Hop Retrieval Support

​

In Traces View:

• Hop count badge on each trace (e.g., “3 hops”)
• Subquery column showing per-hop recall
• Depth indicator for complex trace structures

​

In Detail Views:

• Recall progression chart visualizing improvement across hops
• Delta indicators showing how each hop changed overall recall
• Effectiveness badges for hops that improved recall
• Final context marker showing which hop’s output was used

​

Example Journey

Query: "What is the capital of France and its population?"
├─ retrieval_hop_1: "capital of France" → 72% recall (+72%)
├─ retrieval_hop_2: "population of Paris" → 85% recall (+13%)
└─ context_join: Combined context → 92% recall (+7%) FINAL

​

Working with Logs

​

Debugging a Low-Quality Query

1. Filter by Quality: Poor
2. Click a record to open the detail drawer
3. Check Context Passages:
   • Are relevant docs missing entirely?
   • Are docs present but scored low?
   • Are irrelevant docs ranked high?
4. Check Ground Truth (if available):
   • Which gold docs are missing?
5. Check Seer Evaluation Output:
   • Which requirements were not covered?

​

Investigating a Failed Trace

1. Switch to Traces view
2. Filter by Status: Failed
3. Expand the trace to see which span failed
4. Open the Record Detail for the failed span
5. Check the Error Details section

​

Comparing Multi-Hop Performance

1. In Traces view, toggle to Avg metric mode
2. Sort by Overall ascending to find traces where avg is much lower than final
3. These traces have inefficient early hops that later hops had to compensate for

​

Pagination

• Page size: 10, 25, 50, 100 records per page
• Navigation: Previous/Next with page number display
• URL persistence: Filters and page state are saved in the URL for sharing

​

Coming Soon

• Export: Download filtered logs as CSV/JSON
• Saved filters: Save common filter combinations
• Direct linking: Deep links from monitoring dashboard to filtered logs
• Annotations: Add tags and notes to specific records

​

See Also

• Metrics
• Multi-Hop Retrieval
• Production Monitoring
• Context and Event Schema