Which Model Should You Pick?
Long-context models recommended? Yes — but which one, when, and why. A hands-on guide to balancing cost, speed, and quality.
The Question Everyone Asks
“Which model should I use?”
It’s the most common question after “How do I install this?” The answer is: it depends on what you’re doing.
For Ingestion: Speed and Context
Ingestion is the most token-intensive operation. Two things matter:
- Context window — Your entire Wiki needs to fit
- Cost per token — You’ll run this frequently
Recommended:
- DeepSeek V4-Flash — Lowest cost at $0.14/M tokens. Ideal for batch ingestion.
- Gemini-3.5-Flash — 4× faster output than GPT-5.5.
For Query: Quality Over Speed
Query operations are less token-intensive. Answer quality matters more than speed.
Recommended:
- Claude Sonnet 4.6 — Great quality/cost balance.
- Claude Opus 4.7 — Ultimate quality, use selectively.
Practical Tips
Start with DeepSeek for ingestion, switch to Claude for query. Best of both worlds.
Use Ollama for query, not ingestion. Local models have smaller context windows (8K–128K) — fine for querying, not for processing large sources.
Watch for rate limits. HTTP 429 errors → lower concurrency to 1–2, increase batch delay to 500–800ms.