Performance tuning

Xiajiao itself is lightweight; almost all perceived latency comes from LLM APIs. This page focuses on end-to-end responsiveness.

TIP

Numbers below are indicative. Measure on your own hardware and workload.

LLM calls

A single GPT-4o round trip is often 2–15 seconds.

1. Pick the right model

Use case	Suggested models	Typical latency	Cost
Everyday chat	GPT-4o-mini / Qwen Turbo	1–3s	Very low
Hard reasoning	GPT-4o / Claude Sonnet	3–8s	Medium
Code	Claude Sonnet / DeepSeek Coder	3–10s	Medium
Value	DeepSeek Chat / Qwen Plus	2–5s	Low
Free local	Ollama	Hardware-bound	Free

Not every agent needs the largest model—use small models for rewrite/format tasks.

2. Trim the system prompt

System prompt = SOUL.md + injected memory + RAG hits.

Length drives:

• Time to first token — longer prompts cost more prefill
• Billable tokens
• Remaining context for the conversation

Lever	Action
SOUL.md	Aim for ~500–1000 words of persona/rules
Memory	Inject only top-5 most relevant
RAG	Often top-3 chunks suffice
History	System truncates, but busy channels grow fast

3. Local models (Ollama)

Model	VRAM	CPU	Approx. tok/s
7B	8 GB GPU	8 cores	~30
13B	16 GB	8 cores	~20
33B	24 GB	16 cores	~10
70B	48 GB	16 cores	~5

time ollama run llama3 "Hello world" --verbose

SQLite

WAL

WAL (write-ahead logging) is required for good concurrency:

sqlite3 data/xiajiao.db "PRAGMA journal_mode;"
# expect: wal

Maintenance

sqlite3 data/xiajiao.db "VACUUM;"
sqlite3 data/xiajiao.db "ANALYZE;"
sqlite3 data/xiajiao.db "PRAGMA integrity_check;"

Past ~100 MB DB size, consider monthly VACUUM.

Table sizes

sqlite3 data/xiajiao.db "
SELECT name, SUM(pgsize) AS size_bytes,
  ROUND(SUM(pgsize)/1024.0/1024.0, 2) AS size_mb
FROM dbstat GROUP BY name ORDER BY size_bytes DESC;
"

Node.js memory

Node.js memory monitoring

node -e "
const m = process.memoryUsage();
console.log({
  heapUsed: (m.heapUsed/1024/1024).toFixed(1) + ' MB',
  heapTotal: (m.heapTotal/1024/1024).toFixed(1) + ' MB',
  rss: (m.rss/1024/1024).toFixed(1) + ' MB'
});
"

Memory leak troubleshooting

If RSS keeps climbing:

pm2 monit
pm2 start server/index.js --max-memory-restart 500M

Production

PM2 example

// ecosystem.config.js
module.exports = {
  apps: [{
    name: 'xiajiao',
    script: 'server/index.js',
    instances: 1,
    max_memory_restart: '500M',
    env: {
      NODE_ENV: 'production',
      PORT: 18800,
      OWNER_KEY: 'your-strong-password'
    },
    log_date_format: 'YYYY-MM-DD HH:mm:ss',
    merge_logs: true,
    error_file: 'logs/error.log',
    out_file: 'logs/out.log'
  }]
};

Nginx static cache

location ~* \.(js|css|png|jpg|gif|ico|svg|woff2)$ {
    expires 7d;
    add_header Cache-Control "public, immutable";
}

location /api/ {
    proxy_pass http://127.0.0.1:18800;
    proxy_cache off;
}

Gzip

gzip on;
gzip_types text/plain text/css application/json application/javascript text/xml;
gzip_min_length 1000;
gzip_comp_level 6;

Collaboration chains

Agents run serially. Total time ≈ sum of each hop. Use faster/cheaper models on non-critical steps (editing, translation).

RAG

Indexing

Tip	Effect
Keep docs < ~50 pages / < ~100 KB text	Faster indexing
Remove stale docs	Less noise
Prefer TXT/MD over scanned PDF	Cleaner chunks

Retrieval knobs

BM25 vs vector weight → default 0.5 / 0.5 is fine
RRF k → default 60
top_k → 3 for most cases; 5 for long docs

Large libraries (100+ docs)

sqlite3 data/workspace-xxx/rag.db "INSERT INTO docs_fts(docs_fts) VALUES('rebuild');"
sqlite3 data/workspace-xxx/rag.db "VACUUM;"

(Adjust table names to match your build if they differ.)

Load expectations

Single-process SQLite fits individuals and small teams.

TIP

Concurrency is usually capped by LLM rate limits and latency, not Xiajiao. WAL easily keeps up with human-paced chat.

Quick bench

for i in $(seq 1 100); do
  curl -s -o /dev/null -w "%{time_total}\n" http://localhost:18800/
done | awk '{sum+=$1} END {print "avg:", sum/NR, "s"}'

curl -s -o /dev/null -w "%{time_total}s\n" \
  -H "Cookie: session=your-session-token" \
  http://localhost:18800/api/channels

Monitoring

Signal	Tool	Threshold	Action
Memory	PM2 / htop	> 500 MB	Restart policy
CPU	PM2 / htop	> 80% for 5 min	Inspect loops
Disk	df -h	< 1 GB free	Logs + VACUUM
Main DB	ls -lh data/xiajiao.db	> 1 GB	Archive + VACUUM
Process	PM2 / systemd	Exits	Auto-restart
WAL	ls -lh data/xiajiao.db-wal	> 100 MB	Restart to checkpoint

PM2 monitoring script

# Create a simple monitoring script
cat > /opt/monitor-xiajiao.sh << 'SCRIPT'
#!/bin/bash
MEM=$(pm2 jlist | python3 -c "import sys,json; print(json.load(sys.stdin)[0]['monit']['memory']//1024//1024)")
DISK=$(df -h / | awk 'NR==2{print $5}' | tr -d '%')
DB_SIZE=$(stat -f%z data/xiajiao.db 2>/dev/null || stat -c%s data/xiajiao.db)
DB_MB=$((DB_SIZE/1024/1024))

echo "Memory: ${MEM}MB | Disk: ${DISK}% | DB: ${DB_MB}MB"

[ "$MEM" -gt 500 ] && echo "WARNING: Memory > 500MB"
[ "$DISK" -gt 90 ] && echo "WARNING: Disk > 90%"
[ "$DB_MB" -gt 1024 ] && echo "WARNING: DB > 1GB"
SCRIPT
chmod +x /opt/monitor-xiajiao.sh

Related docs

• Cloud deployment
• Docker deployment
• Security
• Architecture
• Troubleshooting
• Glossary