Insight

Monitor and debug your deployed agents using runtime logs, endpoint logs, and resource metrics (CPU/RAM). All Insight operations use the Runtime API.

---

Overview

AgentBase Insight provides three read-only data sources:

Data Source	Description	API
Runtime Logs	Container stdout/stderr from all replicas	POST /agent-runtimes/{id}/logs
Endpoint Logs	Logs scoped to a specific endpoint	POST /agent-runtimes/{id}/endpoints/{endpointId}/logs
Metrics	Point-in-time CPU and RAM usage	GET /agent-runtimes/{id}/endpoints/{endpointId}/metrics

---

Prerequisites

You need the runtime ID and (for endpoint logs/metrics) the endpoint ID.

TOKEN=$(bash .claude/skills/agentbase/scripts/get_token.sh)

# List runtimes to get IDs
curl -s "https://agentbase.api.vngcloud.vn/runtime/agent-runtimes?page=1&size=20" \
  -H "Authorization: Bearer $TOKEN" | jq '.listData[] | {id, name, status}'

# Get endpoints for a runtime
curl -s "https://agentbase.api.vngcloud.vn/runtime/agent-runtimes/$RUNTIME_ID/endpoints?page=1&size=10" \
  -H "Authorization: Bearer $TOKEN" | jq '.listData[] | {id, name, url, status}'

---

Runtime Logs

Fetch container logs from all replicas of a runtime. Uses offset-based pagination (from/limit).

Portal (GUI)

1. Open https://aiplatform.console.vngcloud.vn/runtime

2. Open the runtime detail page → "Monitor" tab → click an endpoint → "Log" section

1774581394627

---

RESTful API

Prerequisite: All API examples below use $TOKEN — an IAM bearer token. See Configure Authentication for how to obtain it.

RUNTIME_ID="<your-runtime-id>"

# Fetch first 100 log lines
curl -s -X POST "https://agentbase.api.vngcloud.vn/runtime/agent-runtimes/$RUNTIME_ID/logs" \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"from": 0, "limit": 100}' | jq .

Response:

{
  "totalCount": 342,
  "logs": [
    "2026-03-18T09:01:00Z INFO  Starting agent on port 8080",
    "2026-03-18T09:01:02Z INFO  Health check passed",
    "2026-03-18T09:05:33Z INFO  Received request"
  ]
}

Paginate with from:

curl -s -X POST "https://agentbase.api.vngcloud.vn/runtime/agent-runtimes/$RUNTIME_ID/logs" \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"from": 100, "limit": 100}' | jq .

Fetch all logs:

OFFSET=0; LIMIT=500
while true; do
  RESULT=$(curl -s -X POST "https://agentbase.api.vngcloud.vn/runtime/agent-runtimes/$RUNTIME_ID/logs" \
    -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
    -d "{\"from\": $OFFSET, \"limit\": $LIMIT}")
  TOTAL=$(echo "$RESULT" | jq '.totalCount')
  echo "$RESULT" | jq -r '.logs[]'
  OFFSET=$((OFFSET + LIMIT))
  [ "$OFFSET" -ge "$TOTAL" ] && break
done

Filter errors locally:

curl -s -X POST "https://agentbase.api.vngcloud.vn/runtime/agent-runtimes/$RUNTIME_ID/logs" \
  -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  -d '{"from": 0, "limit": 1000}' | jq -r '.logs[]' | grep -i "error\|traceback\|exception\|failed"

Limits: from max = 5000, limit max = 1000

---

Resource Metrics (CPU / RAM)

Get point-in-time CPU and RAM usage for a specific endpoint.

Portal (GUI)

Open the runtime detail page → "Monitor" tab → click an endpoint → "Metrics" section

1774581238335

---

RESTful API

curl -s "https://agentbase.api.vngcloud.vn/runtime/agent-runtimes/$RUNTIME_ID/endpoints/$ENDPOINT_ID/metrics" \
  -H "Authorization: Bearer $TOKEN" | jq .

Response fields: cpuCores (double), memoryBytes (int64)

{
  "cpuCores": 0.42,
  "memoryBytes": 536870912
}

Convert to human-readable:

curl -s "https://agentbase.api.vngcloud.vn/runtime/agent-runtimes/$RUNTIME_ID/endpoints/$ENDPOINT_ID/metrics" \
  -H "Authorization: Bearer $TOKEN" | jq '{
    cpu_cores: .cpuCores,
    memory_mb: (.memoryBytes / 1048576 | floor),
    memory_gb: (.memoryBytes / 1073741824 | . * 100 | floor / 100)
  }'

---

Pseudo-Tailing (Poll Pattern)

Log streaming is not supported — use polling to approximate tailing:

OFFSET=0
while true; do
  RESULT=$(curl -s -X POST "https://agentbase.api.vngcloud.vn/runtime/agent-runtimes/$RUNTIME_ID/logs" \
    -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
    -d "{\"from\": $OFFSET, \"limit\": 50}")
  COUNT=$(echo "$RESULT" | jq '.totalCount')
  echo "$RESULT" | jq -r '.logs[]'
  OFFSET=$COUNT
  sleep 5
done

Caution: Frequent polling generates many API calls. Avoid polling for extended periods.

---

Log Analysis Guide

Common Error Signatures

Pattern	Meaning	Next Step
Traceback (most recent call last)	Python exception	Read the last line for the actual error
ModuleNotFoundError: No module named '...'	Missing dependency	Add to requirements.txt and rebuild image
ImportError: cannot import name '...'	Wrong package version	Check package version compatibility
ConnectionRefusedError / ConnectionError	Cannot reach external service	Verify service URL and auth credentials
401 Unauthorized / 403 Forbidden	Auth failure	Check IAM token and outbound auth config
OSError: [Errno 98] Address already in use	Port 8080 conflict	Ensure only one process binds port 8080
MemoryError / Killed	Out of memory	Upgrade flavor or optimize memory usage
TimeoutError / ReadTimeout	External API timed out	Increase timeout, check LLM endpoint health
Health check failed	/health not returning 200	Fix health endpoint

---

Correlating Logs with Metrics

CPU	RAM	Diagnosis
High	Normal	CPU-bound workload — scale up or optimize
Normal	High	Memory leak or large data structures — scale up or fix leak
Both high	—	Resource exhaustion — scale up flavor
Both low	—	External bottleneck (LLM API latency, network) — add request timing in logs

---

What's Supported

• Log time range filter — Filter logs by a specific time window (start/end timestamp), so you can narrow down exactly when an issue occurred without fetching the entire log history.

• Log keyword search — Search logs by keyword or phrase directly in the query, returning only matching entries without needing to grep locally after fetching.

• Historical metrics — Query CPU and RAM usage over a time range, not just the current point-in-time snapshot. Useful for spotting resource trends, spikes, and patterns leading up to an incident.

---

Troubleshooting

Error	Cause	Fix
401 Unauthorized	Expired IAM token	Re-obtain token
404 Not Found	Wrong runtime or endpoint ID	Verify IDs with list operations
Empty logs	Container never started	Check runtime status; verify image pull succeeded
Logs cut off at 5000	Max offset limit reached	Logs older than 5000 entries are not accessible

---