langsmith-qa-postmortem

Q&A Postmortem — LangSmith Investigation

Investigate EasyOref Q&A graph failures by analyzing LangSmith traces.

When to Use

• User asked the bot a question and got no answer, wrong answer, or timed out
• Intent was misclassified (e.g., security question classified as off_topic)
• Context was empty despite active session / recent attacks
• Answer was hallucinated or in wrong language

Prerequisites

• LangSmith MCP tools (mcp_langsmith_fetch_runs)
• LangSmith project name: easyoref

Procedure

Step 1: Find the Q&A Traces

Q&A traces are also in the easyoref project but have different input shape. Filter by name or search for userMessage:

mcp_langsmith_fetch_runs(
  project_name="easyoref",
  limit=20,
  is_root="true",
  order_by="-start_time",
  preview_chars=100,
  max_chars_per_page=15000,
  filter='search("userMessage")'
)

Alternatively, if you know the approximate time:

filter='and(gt(start_time, "2026-04-09T13:00:00Z"), lt(start_time, "2026-04-09T14:00:00Z"))'

Key input fields for Q&A traces:

• inputs.userMessage: The user's question text
• inputs.chatId: Telegram chat ID
• inputs.language: ru / en / he
• inputs.intent: Classified intent (may be in outputs)

Step 2: Map the Q&A Flow

The Q&A graph has 3 nodes:

Fetch child runs for the trace:

mcp_langsmith_fetch_runs(
  project_name="easyoref",
  trace_id="<trace_id>",
  limit=20,
  order_by="start_time",
  preview_chars=200,
  max_chars_per_page=20000
)

Build a timing table:

Step 3: Diagnose Intent Classification

Check intent-classify output for the intent field.

Valid intents:

• current_alert — questions about active/recent alerts
• recent_history — questions about past alerts (yesterday, last week)
• general_security — general security situation questions
• bot_help — questions about the bot itself
• off_topic — non-security questions (short-circuited, no LLM call)

Common misclassifications:

• Security question classified as off_topic → check intent.ts regex patterns
• History question classified as current_alert → may get empty context if no active session

Step 4: Diagnose Context Gathering

Check context-gather output for the context field.

5 data sources checked (in order):

1. Active Redis session (getActiveSession())
2. Enrichment cache (getSynthesizedInsights())
3. Current Oref API (fetchTzevaAdom())
4. Oref history API (fetchTzevaAdomHistory())
5. Channel posts from Redis (GramJS stored posts)

Failure patterns:

• "No active alert at the moment." only → sources 2-5 not queried (old bug, fixed v2.0.4)
• Empty context despite recent attack → Redis session expired (TTL), check phaseTimeoutMs
• Oref API timeout → fetchTzevaAdomHistory failed logged, context degraded

Step 5: Diagnose Answer Generation

Check answer-generate node and its child ChatOpenRouter LLM calls.

Failure patterns:

• Timeout (>30s): AbortSignal.timeout(30_000) should trigger. If not present → missing timeout fix
• Structured output hung: withStructuredOutput() on some models hangs indefinitely → check if fallback triggered
• Wrong language: Check language in state vs answer text
• No citations: System prompt should instruct [[channel_name]](url) format

LLM call details:

mcp_langsmith_fetch_runs(
  project_name="easyoref",
  trace_id="<trace_id>",
  run_type="llm",
  limit=10,
  preview_chars=300
)

Step 6: Check Rate Limiting

If user reports "no response at all", rate limiter may have blocked (5 questions/min per chatId). This is NOT visible in LangSmith — check RPi logs with the rpi-qa-logs skill.

Known Q&A Bug Patterns

Tips

• Q&A traces are smaller than enrichment traces (usually 3-5 runs total)
• intent-classify and context-gather use zero tokens — if problems are there, it's logic bugs not LLM
• answer-generate is the only LLM node — check for model, latency, structured output issues
• Status callbacks ("🔎 Checking alerts...") are NOT logged to LangSmith — check RPi logs