Performance Optimization

Performance Optimization

Strategies and techniques for maximizing speed and efficiency in documentation discovery.

Core Principles

0. Use context7.com for Instant llms.txt Access

Fastest Approach:

Direct URL construction instead of searching:

Traditional: WebSearch (15-30s) → WebFetch (5-10s) = 20-40s
context7.com: Direct WebFetch (5-10s) = 5-10s

Speed improvement: 2-4x faster

Benefits:

• No search required (instant URL construction)
• Consistent URL patterns
• Reliable availability
• Topic filtering for targeted results

Examples:

GitHub repo:
https://context7.com/vercel/next.js/llms.txt
→ Instant, no search needed

Website:
https://context7.com/websites/imgix/llms.txt
→ Instant, no search needed

Topic-specific:
https://context7.com/shadcn-ui/ui/llms.txt?topic=date
→ Filtered results, even faster

Performance Impact:

Without context7.com:
1. WebSearch for llms.txt: 15s
2. WebFetch llms.txt: 5s
3. Launch agents: 5s
Total: 25s

With context7.com:
1. Direct WebFetch: 5s
2. Launch agents: 5s
Total: 10s (2.5x faster!)

With context7.com + topic:
1. Direct WebFetch (filtered): 3s
2. Process focused results: 2s
Total: 5s (5x faster!)

1. Minimize Sequential Operations

The Problem:

Sequential operations add up linearly:

Total Time = Op1 + Op2 + Op3 + ... + OpN

Example:

Fetch URL 1: 5 seconds
Fetch URL 2: 5 seconds
Fetch URL 3: 5 seconds
Total: 15 seconds

The Solution:

Parallel operations complete in max time of slowest:

Total Time = max(Op1, Op2, Op3, ..., OpN)

Example:

Launch 3 agents simultaneously
All complete in: ~5 seconds
Total: 5 seconds (3x faster!)

2. Batch Related Operations

Benefits:

• Fewer context switches
• Better resource utilization
• Easier to track
• More efficient aggregation

Grouping Strategies:

By topic:

Agent 1: All authentication-related docs
Agent 2: All database-related docs
Agent 3: All API-related docs

By content type:

Agent 1: All tutorials
Agent 2: All reference docs
Agent 3: All examples

By priority:

Phase 1 (critical): Getting started, installation, core concepts
Phase 2 (important): Guides, API reference, configuration
Phase 3 (optional): Advanced topics, internals, optimization

3. Smart Caching

What to cache:

• Repomix output (expensive to generate)
• llms.txt content (static)
• Repository structure (rarely changes)
• Documentation URLs (reference list)

When to refresh:

• User requests specific version
• Documentation updated (check last-modified)
• Cache older than session
• User explicitly requests fresh data

4. Early Termination

When to stop:

✓ User's core needs met
✓ Critical information found
✓ Time limit approaching
✓ Diminishing returns (90% coverage achieved)

How to decide:

After Phase 1 (critical docs):
- Review what was found
- Check against user request
- If 80%+ covered → deliver now
- Offer to fetch more if needed

Performance Patterns

Pattern 1: Parallel Exploration

Scenario: llms.txt contains 10 URLs

Slow approach (sequential):

Time: 10 URLs × 5 seconds = 50 seconds

Step 1: Fetch URL 1 (5s)
Step 2: Fetch URL 2 (5s)
Step 3: Fetch URL 3 (5s)
...
Step 10: Fetch URL 10 (5s)

Fast approach (parallel):

Time: ~5-10 seconds total

Step 1: Launch 5 Explorer agents (simultaneous)
  Agent 1: URLs 1-2
  Agent 2: URLs 3-4
  Agent 3: URLs 5-6
  Agent 4: URLs 7-8
  Agent 5: URLs 9-10

Step 2: Wait for all (max time: ~5-10s)
Step 3: Aggregate results

Speedup: 5-10x faster

Pattern 2: Lazy Loading

Scenario: Documentation has 30+ pages

Slow approach (fetch everything):

Time: 30 URLs × 5 seconds ÷ 5 agents = 30 seconds

Fetch all 30 pages upfront
User only needs 5 of them
Wasted: 25 pages × 5 seconds ÷ 5 = 25 seconds

Fast approach (priority loading):

Time: 10 URLs × 5 seconds ÷ 5 agents = 10 seconds

Phase 1: Fetch critical 10 pages
Review: Does this cover user's needs?
If yes: Stop here (saved 20 seconds)
If no: Fetch additional as needed

Speedup: Up to 3x faster for typical use cases

Pattern 3: Smart Fallbacks

Scenario: llms.txt not found

Slow approach (exhaustive search):

Time: ~5 minutes

Try: docs.library.com/llms.txt (30s timeout)
Try: library.dev/llms.txt (30s timeout)
Try: library.io/llms.txt (30s timeout)
Try: library.org/llms.txt (30s timeout)
Try: www.library.com/llms.txt (30s timeout)
Then: Fall back to repository

Fast approach (quick fallback):

Time: ~1 minute

Try: docs.library.com/llms.txt (15s)
Try: library.dev/llms.txt (15s)
Not found → Immediately try repository (30s)

Speedup: 5x faster

Pattern 4: Incremental Results

Scenario: Large documentation set

Slow approach (all-or-nothing):

Time: 5 minutes until first result

Fetch all documentation
Aggregate everything
Present complete report
User waits 5 minutes

Fast approach (streaming):

Time: 30 seconds to first result

Phase 1: Fetch critical docs (30s)
Present: Initial findings
Phase 2: Fetch important docs (60s)
Update: Additional findings
Phase 3: Fetch supplementary (90s)
Final: Complete report

Benefit: User gets value immediately, can stop early if satisfied

Optimization Techniques

Technique 1: Workload Balancing

Problem: Uneven distribution causes bottlenecks

Bad distribution:
Agent 1: 1 URL (small) → finishes in 5s
Agent 2: 10 URLs (large) → finishes in 50s
Total: 50s (bottlenecked by Agent 2)

Solution: Balance by estimated size

Good distribution:
Agent 1: 3 URLs (medium pages) → ~15s
Agent 2: 3 URLs (medium pages) → ~15s
Agent 3: 3 URLs (medium pages) → ~15s
Agent 4: 1 URL (large page) → ~15s
Total: ~15s (balanced)

Technique 2: Request Coalescing

Problem: Redundant requests slow things down

Bad:
Agent 1: Fetch README.md
Agent 2: Fetch README.md (duplicate!)
Agent 3: Fetch README.md (duplicate!)
Wasted: 2 redundant fetches

Solution: Deduplicate before fetching

Good:
Pre-processing: Identify unique URLs
Agent 1: Fetch README.md (once)
Agent 2: Fetch INSTALL.md
Agent 3: Fetch API.md
Share: README.md content across agents if needed

Technique 3: Timeout Tuning

Problem: Default timeouts too conservative

Slow:
WebFetch timeout: 120s (too long for fast sites)
If site is down: Wait 120s before failing

Solution: Adaptive timeouts

Fast:
Known fast sites (official docs): 30s timeout
Unknown sites: 60s timeout
Large repos: 120s timeout
If timeout hit: Immediately try alternative

Technique 4: Selective Fetching

Problem: Fetching irrelevant content

Wasteful:
Fetch: Installation guide ✓ (needed)
Fetch: API reference ✓ (needed)
Fetch: Internal architecture ✗ (not needed for basic usage)
Fetch: Contributing guide ✗ (not needed)
Fetch: Changelog ✗ (not needed)

Solution: Filter by user needs

Efficient:
User need: "How to get started"
Fetch only: Installation, basic usage, examples
Skip: Advanced topics, internals, contribution
Speedup: 50% less fetching

Performance Benchmarks

Target Times

Real-World Examples

Fast case (Next.js with llms.txt):

00:00 - Start
00:05 - Found llms.txt
00:10 - Fetched content (12 URLs)
00:15 - Launched 4 agents
00:45 - All agents complete
00:55 - Report ready
Total: 55 seconds ✓

Medium case (Repository without llms.txt):

00:00 - Start
00:15 - llms.txt not found
00:20 - Found repository
00:30 - Cloned repository
02:00 - Repomix complete
02:30 - Analyzed output
02:45 - Report ready
Total: 2m 45s ✓

Slow case (Scattered documentation):

00:00 - Start
00:30 - llms.txt not found
00:45 - Repository not found
01:00 - Launched 4 Researcher agents
05:00 - All research complete
06:00 - Aggregated findings
06:30 - Report ready
Total: 6m 30s (acceptable for research)

Common Performance Issues

Issue 1: Too Many Agents

Symptom: Slower than sequential

Problem:
Launched 15 agents for 15 URLs
Overhead: Agent initialization, coordination
Result: Slower than 5 agents with 3 URLs each

Solution:

Max 7 agents per batch
Group URLs sensibly
Use phases for large sets

Issue 2: Blocking Operations

Symptom: Agents waiting unnecessarily

Problem:
Agent 1: Fetch URL, wait for Agent 2
Agent 2: Fetch URL, wait for Agent 3
Agent 3: Fetch URL
Result: Sequential instead of parallel

Solution:

Launch all agents independently
No dependencies between agents
Aggregate after all complete

Issue 3: Redundant Fetching

Symptom: Same content fetched multiple times

Problem:
Phase 1: Fetch installation guide
Phase 2: Fetch installation guide again
Result: Wasted time

Solution:

Cache fetched content
Check cache before fetching
Reuse within session

Issue 4: Late Bailout

Symptom: Continuing when should stop

Problem:
Found 90% of needed info after 1 minute
Spent 4 more minutes on remaining 10%
Result: 5x time for marginal gain

Solution:

Check progress after critical phase
If 80%+ covered → offer to stop
Only continue if user wants comprehensive

Performance Monitoring

Key Metrics

Track these times:

- llms.txt discovery: Target <30s
- Repository clone: Target <60s
- Repomix processing: Target <2min
- Agent exploration: Target <60s
- Total time: Target <3min for typical case

Performance Report Template

## Performance Summary

**Total time**: 1m 25s
**Method**: llms.txt + parallel exploration

**Breakdown**:
- Discovery: 15s (llms.txt search & fetch)
- Exploration: 50s (4 agents, 12 URLs)
- Aggregation: 20s (synthesis & formatting)

**Efficiency**: 8.5x faster than sequential
(12 URLs × 5s = 60s sequential, actual: 50s parallel)

When to Optimize Further

Optimize if:

• Total time >2x target
• User explicitly requests "fast"
• Repeated similar queries (cache benefit)
• Large documentation set (>20 URLs)

Don't over-optimize if:

• Already meeting targets
• One-time query
• User values completeness over speed
• Research requires thoroughness

Quick Optimization Checklist

Before Starting

• Check if content already cached
• Identify fastest method for this case
• Plan for parallel execution
• Set appropriate timeouts

During Execution

• Launch agents in parallel (not sequential)
• Use single message for multiple agents
• Monitor for bottlenecks
• Be ready to terminate early

After First Phase

• Assess coverage achieved
• Determine if user needs met
• Decide: continue or deliver now
• Cache results for potential reuse

Optimization Decision Tree

Need documentation?
  ↓
Check cache
  ↓
HIT → Use cached (0s) ✓
MISS → Continue
  ↓
llms.txt available?
  ↓
YES → Parallel agents (30-60s) ✓
NO → Continue
  ↓
Repository available?
  ↓
YES → Repomix (2-5min)
NO → Research (3-7min)
  ↓
After Phase 1:
80%+ coverage?
  ↓
YES → Deliver now (save time) ✓
NO → Continue to Phase 2