Debug a Bad Answer¶

When Lattice returns an unhelpful or incorrect answer, the issue is usually in selection (wrong atoms retrieved) or synthesis (model confabulated). This guide walks through how to diagnose which.

Step 1: Check what atoms were returned¶

The web UI shows source chips below every answer. Click each chip to see the atom's content and metadata. If the atoms look irrelevant, the problem is in selection. If the atoms look correct but the answer doesn't match them, the problem is in synthesis.

Step 2: Use the feedback buttons¶

Click 👎 on a bad answer. If specific atom chips were wrong sources, click the ✕ on those chips to dismiss them. This feeds the quality score update loop — dismissed atoms score lower in future recalls.

Step 3: Check the query reformulation¶

If you asked a follow-up question, Lattice may have reformulated it. The web UI shows the reformulated query in small text below your question. If reformulation distorted your intent, this is the likely culprit.

To disable reformulation for a query, start a fresh conversation (clear button) rather than following up.

Step 4: Check the selection directly¶

Use the lattice_select MCP tool (or POST /api/query with a debug client) to inspect the raw atom pack:

curl -s -X POST http://localhost:7337/api/answer \
  -H "Content-Type: application/json" \
  -d '{"question": "your question here", "conversation_history": []}' \
  | python3 -m json.tool

Look at the atoms array — are the right atoms present?

Step 5: Check for supersession issues¶

If you recently updated a fact, the old atom may be showing up. Check:

grep -r "subject: <your subject>" ~/.lattice/*.md

Any atom with is_superseded: true should not appear in answers. If it does, run:

uv run lattice graph rebuild

Step 6: Enable tracing¶

export LATTICE_TRACE=true
uv run lattice-daemon

Re-run the query. Check ~/.lattice/traces.jsonl for the trace of that query — it shows BM25 seed count, dense seed count, BFS expansion count, final atom count, and synthesis latency.

Step 7: Check dense search¶

If you have LATTICE_DENSE_SEEDS=1, the dense index might be surfacing irrelevant atoms. Try disabling it temporarily:

LATTICE_DENSE_SEEDS=0 uv run lattice-daemon

If the answer improves, the dense model may be finding spurious vocabulary matches.

Common patterns¶

Symptom	Likely cause	Fix
Answer says "I don't have information" but you definitely captured it	BM25 vocabulary mismatch	Enable `LATTICE_DENSE_SEEDS=1`
Answer is about the wrong topic	Reformulation distorted the query	Clear conversation history and ask directly
Answer cites old superseded facts	Supersession didn't run properly	`lattice graph rebuild`
Answer is correct but doesn't cite the most relevant atoms	Quality scores not calibrated yet	Give 👍 feedback on good answers
Synthesis confabulates details not in atoms	Model is hallucinating	Use a larger/better synthesis model (`SYNTHESIS_MODEL`)