Advanced usage
Use Reticuli more deliberately when retrieval, fine-tuning, models, and deployment choices matter.
This page is for users who already understand the basic workflow and want better control.
Advanced usage is mostly discipline: change one thing, retest same prompts, compare outputs, then promote better version.
Retrieval versus fine-tuning
Use retrieval for information. Use fine-tuning for behavior.
| Need | Best starting point |
|---|---|
| Keep answers current as docs change | Knowledge base |
| Cite or inspect source passages | Knowledge base |
| Match team tone | Fine-tuning |
| Follow a strict answer structure | Fine-tuning |
| Improve factual accuracy and answer style | Both |
Split projects by job
Do not force one assistant to serve unrelated audiences.
Split projects when:
- different users need different source access
- the tone or format is different
- policies conflict between teams
- testing becomes too broad to interpret
Version thinking
Treat major changes as versions:
- Save the current prompt, source set, and examples.
- Change one important thing.
- Retest the same prompt set.
- Compare outputs.
- Promote the better version.
Source quality hierarchy
Use this priority order:
- Current official policy or docs.
- Approved examples from domain experts.
- Recent support or workflow history.
- Draft notes with clear owner review.
- Old or imported material only after cleanup.
Operational ownership
Every production assistant needs owners for:
- source freshness
- training examples
- prompt and behavior changes
- deployment keys
- billing and usage review
- launch and rollback decisions
Change management
Use one change per review cycle:
| Change | Retest focus |
|---|---|
| Add source files | Factual answers and citations |
| Remove stale source | Old policy no longer appears |
| Update assistant brief | Refusals, tone, and structure |
| Add examples | Format and judgment consistency |
| New fine-tune | Regression against full prompt set |
| New API key | Integration errors and usage spike |
Red flags
Investigate before launch if:
- the assistant gives confident answers without sources
- a small source change creates broad behavior changes
- users ask questions outside the assistant job
- no one can explain why an answer changed