Skip to main content

Cache Management Best Practices

Learn proven strategies and guidelines for effective cache management that maximize chatbot performance and user satisfaction.

Core Principles​

1. Quality Over Quantity​

Principle: A few high-quality cache entries outperform many low-quality ones.

Why it matters:

  • Accurate entries build user trust
  • Poor entries damage credibility
  • Maintenance scales with quantity

In practice:

  • βœ… Start with 10-20 excellent entries
  • βœ… Expand gradually based on usage data
  • ❌ Don't rush to cache 100+ entries immediately

Example:

  • Bad approach: Cache 50 entries in one day, many with guessed answers
  • Good approach: Cache 10 verified entries, monitor performance, expand based on data

2. User-Centric Design​

Principle: Cache entries should match how users actually ask questions, not how you think they should ask.

Why it matters:

  • Users don't know official terminology
  • Real questions are often casual or fragmented
  • Matching user language = higher hit rate

In practice:

  • βœ… Review actual user questions in Conversations
  • βœ… Add variations matching real user phrasings
  • βœ… Include misspellings and abbreviations
  • ❌ Don't only use official/formal language

Example:

  • User asks: "who's teaching 3000"
  • Bad cache question: "Who is the instructor for HFIM 3000 in Fall 2026?"
  • Good cache question: "Who teaches HFIM 3000?" with variations like "who's teaching 3000", "3000 professor", "instructor for hospitality 3000"

3. Continuous Improvement​

Principle: Cache management is an ongoing process, not a one-time setup.

Why it matters:

  • Information changes (faculty, requirements, dates)
  • User questions evolve
  • Performance metrics reveal optimization opportunities

In practice:

  • βœ… Review entries weekly
  • βœ… Update based on feedback
  • βœ… Monitor metrics and trends
  • ❌ Don't "set and forget"

Recommended Schedule:

  • Daily (5 min): Check for negative feedback
  • Weekly (30 min): Review metrics, update high-traffic entries
  • Monthly (2 hours): Deep audit, update stale entries
  • Quarterly (4 hours): Comprehensive review, strategic planning

Content Guidelines​

Writing Questions​

Be Natural and Conversational​

Good Questions:

  • "What is the HFIM program?"
  • "How do I apply to HFIM?"
  • "Who is the academic advisor?"

Bad Questions:

  • "HFIM program information retrieval query" (too technical)
  • "Question regarding the application process..." (too formal)
  • "Provide advisor contact details" (command, not question)

Use the Most Common Phrasing​

How to find it:

  1. Check Conversations for actual user questions
  2. Use the most frequently asked version
  3. Add other versions as variations

Example:

  • Most common: "What is HFIM?" (main question)
  • Variations: "Tell me about HFIM", "Explain HFIM program", "HFIM info"

Avoid Time-Specific References​

Don't include (unless necessary):

  • "Spring 2026" in the question
  • "Current semester"
  • "This year"

Why: Questions should be timeless. Update the answer, not the question.

Exception: When the question is explicitly about a specific time period.

Good:

  • Question: "When is the HFIM application deadline?"
  • Answer: "The Fall 2027 application deadline is January 15, 2027." (TTL: 7 days)

Writing Responses​

Structure for Clarity​

Use formatting to improve readability:

**Bold**: Emphasize key points
β€’ Bullets: List multiple items
Paragraphs: Separate concepts

Examples:
1. Numbered lists for steps
2. Second step
3. Third step

Example Response:

The HFIM program requires:

β€’ Minimum 3.0 GPA
β€’ SAT 1200+ or ACT 24+
β€’ Two letters of recommendation
β€’ Personal statement (500 words)

**Application deadline**: January 15 for Fall admission

Questions? Email admissions@uga.edu

Be Complete But Concise​

Balance thoroughness with brevity:

  • βœ… Answer the full question
  • βœ… Include relevant details
  • ❌ Don't write essays
  • ❌ Don't oversimplify

Word count guidelines:

  • Simple questions: 30-75 words
  • Moderate questions: 75-150 words
  • Complex questions: 150-300 words
  • Over 300 words: Consider if this should be a document reference instead

Always Include Sources​

Why: Transparency builds trust

Format:

[
{
"filename": "HFIM_Handbook_2026.pdf",
"page": 12,
"section": "Admission Requirements"
}
]

Best practices:

  • βœ… Use actual source documents
  • βœ… Include page numbers
  • βœ… Cite multiple sources when applicable
  • ❌ Don't fabricate sources
  • ❌ Don't leave empty if sources exist

Use Consistent Style​

Maintain consistency across all cache entries:

  • Same tone (professional but friendly)
  • Same formatting patterns
  • Same level of detail
  • Same terminology

Example Style Guide:

- Use "HFIM" (not "H.F.I.M." or "hfim")
- Write "Professor" (not "Prof" unless in variation)
- Include links as: email@uga.edu (not hyperlinks in markdown)
- End with contact info when relevant

Confidence Score Guidelines​

Scoring Framework​

Use this framework to assign confidence scores:

ScoreUse WhenExamples
0.95-1.0Facts verified from official sources, unlikely to changeProgram name, university mission, core requirements
0.85-0.94Current information from reliable sourcesCourse descriptions, faculty bios, current policies
0.70-0.84General advice or information that may varyCareer advice, optional recommendations
Below 0.70Uncertain or needs verificationUnconfirmed details, outdated information, opinions

Confidence Do's and Don'ts​

DO:

  • βœ… Lower confidence for time-sensitive info (use short TTL instead)
  • βœ… Use 0.95+ for permanent facts
  • βœ… Adjust confidence after verification
  • βœ… Be honest about uncertainty

DON'T:

  • ❌ Default everything to 1.0
  • ❌ Use low confidence as a substitute for inactivating entries
  • ❌ Guessβ€”verify first
  • ❌ Use confidence < 0.70 for active entries (inactivate instead)

TTL (Time-To-Live) Guidelines​

Setting Appropriate TTL​

7 Days:

  • Event dates and deadlines
  • Temporary office hours
  • Current semester-specific information

30 Days:

  • Course schedules
  • Faculty office hours
  • Contact information (may change)

90 Days:

  • Degree requirements
  • Program policies
  • Course prerequisites

180+ Days:

  • Historical information
  • Mission statements
  • Foundational concepts

TTL Strategy​

Link TTL to Review Frequency:

TTL = How often you should verify this information is still accurate

Set reminders:

  • Use calendar apps
  • Schedule batch reviews
  • Don't rely on memory

Example Workflow:

Every Monday: Review entries with TTL ≀ 7 days
Every 1st of month: Review entries with TTL ≀ 30 days
Every quarter: Review entries with TTL ≀ 90 days

Variation Best Practices​

Variation Quantity​

Guidelines by Entry Type:

Entry TypeVariationsRationale
High-traffic10-15Maximum coverage for popular questions
Medium-traffic5-10Balanced coverage
Low-traffic3-5Minimal but sufficient
Niche topics3-5Users asking know what to ask

Variation Quality​

Include These Types:

  1. Formal vs. Casual

    • "What are the admission requirements?" (formal)
    • "How do I get in?" (casual)
  2. Complete vs. Fragment

    • "Who is the HFIM academic advisor?" (complete)
    • "HFIM advisor" (fragment)
  3. Abbreviations

    • "HFIM program information"
    • "Hospitality and Food Industry Management program information"
  4. Different Word Orders

    • "When is the application deadline?"
    • "Application deadline when?"
  5. Common Misspellings

    • "professer" for "professor"
    • "recomendation" for "recommendation"
  6. Related Terms

    • "instructor" / "professor" / "teacher"
    • "apply" / "application" / "enroll"

Variation Testing​

After adding variations:

  1. Wait 1-2 weeks for data
  2. Check "Times Served" - did it increase?
  3. Review actual user questions in Conversations
  4. Add missing variations based on real queries
  5. Remove ineffective variations that don't match

Iterate: Continuously refine based on usage.


Status Management​

Active vs. Inactive​

Use Active When:

  • βœ… Information is current and accurate
  • βœ… Entry is ready for users
  • βœ… You're confident in the answer

Use Inactive When:

  • ⏸️ Updating content (temporarily)
  • ⏸️ Seasonal information out of season
  • ⏸️ Waiting for verification
  • ⏸️ Testing new variations before going live

Don't Delete Unless:

  • ❌ Information is permanently irrelevant
  • ❌ Entry was created by mistake
  • ❌ Duplicate of another entry

Why not delete?

  • You might need it later
  • Preserves admin notes and history
  • Can reactivate if information becomes relevant again

Admin Notes Best Practices​

What to Document​

Use Admin Notes for:

  1. Change History

    Updated 1/15/2026: Changed deadline from Jan 10 to Jan 15
    Verified with Dr. Smith
  2. Source Details

    Based on Fall 2026 HFIM Handbook, page 12
    Cross-referenced with admissions website
  3. Review Reminders

    Re-review when Spring 2027 handbook is published
    Check with program coordinator before Fall 2026
  4. Collaboration Notes

    Dr. Johnson requested this specific wording
    Marketing team approved this phrasing
  5. Known Issues

    Users sometimes confuse this with HFIM 4000 requirements
    Consider splitting into two entries

Admin Notes Style​

Keep notes:

  • βœ… Brief and scannable
  • βœ… Dated with ISO format (2026-01-15)
  • βœ… Signed (initials or email)
  • βœ… Action-oriented

Example:

2026-01-15 (JS): Updated admission requirements for Fall 2026.
Verified with admissions office. Re-check in Aug 2026 for Fall 2027.

Maintenance Workflows​

Weekly Maintenance (15-30 minutes)​

Checklist:

  • Check Dashboard metrics
  • Review 5-10 highest-traffic entries for accuracy
  • Check for negative feedback in Conversations
  • Update any time-sensitive entries (deadlines, events)
  • Add variations to underperforming entries

Monthly Maintenance (1-2 hours)​

Checklist:

  • Review all entries with TTL ≀ 30 days
  • Audit entries with "Times Served = 0"
  • Generate variations for new entries
  • Update entries with "Last Updated > 60 days"
  • Review and respond to negative feedback
  • Document metrics in tracking log

Quarterly Maintenance (3-4 hours)​

Checklist:

  • Comprehensive audit of all cache entries
  • Update all entries with "Last Updated > 90 days"
  • Review and refine variations based on user data
  • Analyze metrics trends over the quarter
  • Identify gaps (topics not cached but frequently asked)
  • Plan cache expansion for next quarter
  • Update documentation and style guides

Team Collaboration​

When Multiple Admins Manage Cache​

Assign Ownership by Topic:

Admin A: Admissions & Application
Admin B: Courses & Curriculum
Admin C: Faculty & Staff

Communicate Changes:

  • Use Admin Notes extensively
  • Hold regular team meetings
  • Share weekly metric reports
  • Document decisions

Establish Style Guide:

  • Agree on tone and formatting
  • Use consistent terminology
  • Share examples of good entries
  • Review each other's work

Handoff Best Practices​

When transferring cache management:

  1. Document everything

    • Current state and metrics
    • Ongoing projects
    • Known issues
  2. Schedule training

    • Walk through key entries
    • Explain workflows
    • Answer questions
  3. Shadow period

    • New admin observes for 2 weeks
    • New admin leads with supervision for 2 weeks
    • Independent with check-ins

Common Pitfalls to Avoid​

Pitfall 1: Over-Caching​

Problem: Caching too many entries too quickly

Consequences:

  • Overwhelming maintenance burden
  • Lower quality per entry
  • Difficulty tracking what needs updates

Solution:

  • Start with 10-20 high-impact entries
  • Expand by 5-10 entries per month
  • Prioritize quality

Pitfall 2: Under-Maintaining​

Problem: Creating cache entries but never updating them

Consequences:

  • Outdated information
  • User distrust
  • Negative feedback

Solution:

  • Set up regular review schedule
  • Use TTL as reminders
  • Monitor metrics weekly

Pitfall 3: Ignoring User Feedback​

Problem: Not reviewing negative feedback or actual user questions

Consequences:

  • Cache doesn't match user needs
  • Low hit rate
  • Frustrated users

Solution:

  • Check Conversations weekly
  • Add variations based on real queries
  • Respond to negative feedback quickly

Pitfall 4: Guessing Instead of Verifying​

Problem: Creating cache entries without confirming accuracy

Consequences:

  • Incorrect information
  • Damaged credibility
  • Liability issues

Solution:

  • Always verify with official sources
  • Set low confidence if uncertain
  • Document sources clearly

Pitfall 5: Set-and-Forget Mentality​

Problem: Assuming cache will work indefinitely without maintenance

Consequences:

  • Stale information
  • Poor performance over time
  • Wasted initial effort

Solution:

  • Treat cache as living system
  • Schedule regular reviews
  • Monitor metrics continuously

Measuring Success​

Key Performance Indicators (KPIs)​

Track these metrics:

  1. Cache Hit Rate - Target: 65-75%
  2. Success Rate - Target: 75-85%
  3. Average Times Served - Target: Increasing trend
  4. Entries with Times Served = 0 - Target: < 20%
  5. Average Variations per Entry - Target: 7-12

Qualitative Indicators​

Monitor these signs:

  • βœ… Fewer negative feedback comments
  • βœ… Users finding information quickly
  • βœ… Positive feedback mentions
  • βœ… Lower support ticket volume
  • βœ… Faster average response times

Next Steps​

Implement These Practices​

  1. Start with core principles

    • Focus on quality
    • Match user language
    • Commit to continuous improvement
  2. Set up workflows

    • Schedule weekly/monthly reviews
    • Create metric tracking log
    • Establish team processes (if applicable)
  3. Monitor and adjust

    • Track KPIs monthly
    • Refine based on data
    • Stay flexible

Additional Resources​


Remember: Best practices evolve with your needs. Start with these guidelines, adapt based on your specific context, and always prioritize user experience!