Optimization & Troubleshooting
Goal
This guide helps with two things:
- reducing cost and noise
- knowing what to check first when something goes wrong
If you have not yet validated the main workflow, start with:
1. Do the highest-value optimizations first
The most effective optimizations are usually not more complexity. They are:
- narrowing source scope
- fixing deduplication strategy
- reducing report steps
- using premium models only where they matter
Those four changes solve most cost and stability issues.
2. Source-layer optimization
Too much noise
Check first:
- Are keywords too broad?
- Is the time range too large?
- Do you already have more sources than you actually need?
Too much duplication
Check first:
- Is the dedup strategy wrong?
- Are multiple sources collecting the same material?
- Are you collecting the same source both via RSS and Search?
Slow collection or high failure rate
Check first:
- Does the source depend on heavy rendering?
- Does the site often require cookies or login state?
- Is the schedule too aggressive?
3. Report-layer optimization
Cost is too high
Start with these:
- move summarization steps to cheaper models
- remove steps that do not add real value
- shorten prompts and remove unnecessary context
- pass only the context a step actually needs
Output quality is unstable
Check first:
- Is each step trying to do too much?
- Is the input itself too weak?
- Does the prompt clearly define format and boundaries?
Generation is too slow
Check first:
- Are all steps using heavy models?
- Is the input context too large?
- Is the template repeating work across steps?
4. Ask optimization
Ask cannot find relevant content
Check first:
- Does the library actually contain relevant content?
- Is the question too broad?
- Did recent collection runs succeed?
Ask cost is too high
Check first:
- Are you using general mode too often?
- Are questions too broad, causing large context windows?
- Are you repeating the same question instead of following up?
5. Trigger Inbox optimization
Low match rate
Check first:
- Is external triggering enabled for the template?
- Are template tags clear enough to match common requests?
- Do requests include time range and topic?
Slow processing
Check first:
- Is there no existing report to reuse?
- Is template generation itself slow?
- Have related sources recently failed?
6. A practical debugging order
When you do not know where the problem is, check in this order:
- Did sources collect successfully?
- Does the library contain usable content?
- Can the report template run manually?
- Are Delivery / Ask / Trigger Inbox simply consuming an upstream failure?
Do not start by changing models or rewriting prompts.
Many problems are input-chain problems, not LLM problems.
7. Common cost-control rules
- keep high-frequency sources narrow
- keep report templates small
- reserve premium models for key steps
- keep Delivery email-first by default
- validate Trigger Inbox with email first before adding more channels
8. Where to look when things fail
- FAQ - user-facing troubleshooting
- Credits & Logs - cost and task records
- Reliability - failover and stability behavior
- Atomic Billing - refunds and billing consistency