Logs Explorer

Drill into individual queries to understand exactly what happened—which documents were retrieved, how Seer scored them, and why quality metrics are what they are.

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

---

Why Use Logs?

Aggregate dashboards show trends, but logs answer the hard questions:

Question	Answer in Logs
Why did this query fail?	See exactly which documents were retrieved and why they were marked irrelevant
Which document is missing?	Ground truth validation shows which gold docs weren't retrieved
How did multi-hop perform?	Trace view shows recall progression across each hop
What did Seer actually see?	Full context passages with relevance labels and scores

---

Two View Modes

The Logs page offers two ways to explore your data:

Traces View (Default)

Best for multi-hop RAG and understanding end-to-end query flows.

• 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

Best for individual record inspection and debugging specific retrievals.

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

When to Use Which

• Use Traces when debugging why a multi-step query failed
• Use Records when scanning for low-performing individual retrievals

Traces view with an expanded multi-hop trace showing the span tree, metrics per hop, and final context marker.

---

Filtering & Search

Find specific queries quickly using the filter toolbar:

Filter	Options
Search	Free-text search across task/query content
Status	All, Succeeded only, Failed only
Quality	All, Excellent, Good, Fair, Poor
Environment	Filter by your environment (dev/staging/prod)

Filters work across both Traces and Records views.

---

Traces View Deep Dive

Trace Table Columns

Column	Description
When	Timestamp of the first span in the trace
Query / Task	The original user query
Spans	Number of spans (retrieval hops) in the trace
Overall	Final recall against the main question (toggle Final/Avg)
Subquery	Per-hop recall against hop-specific queries
Depth	Maximum nesting depth of the trace tree
Latency	Total end-to-end latency
Quality	Quality grade based on final recall

Metric Mode Toggle

Switch between two views of the Overall metric:

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

For most debugging, Final is what matters—it's what the user actually experienced.

Expanding a Trace

Click any trace row to expand and see its spans in a tree structure:

• 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

Click the View button (↗) on any trace to open the full trace sidebar:

Section	What You See
Task / Question	Original query with trace ID and span count
Final Answer Quality	Recall, precision, F1, nDCG with visual bars
Journey Timeline	Recall progression chart + timeline of each hop
Ground Truth Validation	GT recall/precision if gold docs were provided
Performance Breakdown	Total latency, average per hop, slowest/fastest spans

Trace Detail Sidebar with Final Answer Quality metrics and Journey Timeline showing recall progression from 50% → 100% across hops.

---

Records View Deep Dive

Records Table Columns

Column	Description
When	Timestamp of the record
Task / Question	The query that was evaluated
Recall	Fraction of requirements covered
Precision	Fraction of documents supporting requirements
F1	Harmonic mean of recall and precision
Latency	Seer evaluation latency
Quality	Grade (Excellent/Good/Fair/Poor)

Sorting

Click any column header to sort:

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

---

Record Detail Drawer

Click any record (in either view) to open the detail drawer:

Sections

Section	Contents
Task / Question	The original query text
Subquery (multi-hop)	The hop-specific query if part of a trace
Evaluation Metrics	Recall, precision, F1, nDCG, latency, quality grade
Overall Contribution (multi-hop)	How this hop contributed to the final answer
Ground Truth	GT recall/precision if gold docs were provided
Context Passages	All retrieved documents with relevance labels
Seer Evaluation Output	The raw XML output from Seer's evaluator
Trace Info	Trace ID and span ID for correlation

Passage Visualization

Each passage in the context shows:

• 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

Color coding:

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

Record Detail Drawer showing a multi-hop span with subquery, metrics, overall contribution, and ground truth validation.

---

Ground Truth Validation

When you log records with gold_doc_ids, the detail views show:

Metric	Formula	What It Tells You
GT Recall	Gold found / Total gold	% of known-good docs you retrieved
GT Precision	Gold found / Total retrieved	% of retrieved docs that are gold
Missing Gold Docs	List of IDs	Exactly which docs your retriever missed

This is crucial for understanding retrieval quality independent of Seer's evaluation.

---

Multi-Hop Retrieval Support

For traces with multiple spans, the Logs page shows the complete journey:

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

Both views support 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 — Metric definitions
• Multi-Hop Retrieval — Logging traces
• Production Monitoring — Aggregate dashboards
• Context and Event Schema — Data format