Models for RAG: answer over your knowledge
RAG = retrieve + generate. Embeddings fetch chunks; chat answers without over-hallucinating. This collection is the generate side: models that follow instructions, cite context, and do not burn budget on every query.
Models in this collection
| Model | Type | Context | Price |
|---|---|---|---|
 GPT-4.1 mini (legacy)OpenAI | Chat | 1M tokens | $0.436/M in · $1.7441/M out |
 Gemini 2.5 FlashGoogle | Chat | 1M tokens | $0.327/M in · $2.7251/M out |
 Claude Haiku 4.5Anthropic | Chat | 200k tokens | $1.0901/M in · $5.4503/M out |
| GPT-5.4 miniOpenAI | Chat | 272k tokens | $0.8175/M in · $4.9053/M out |
| Gemini 2.5 ProGoogle | Chat | 2M tokens | $1.3626/M in · $10.9006/M out |
| Claude Sonnet 5Anthropic | Chat | 1M tokens | $3.2702/M in · $16.3509/M out |
 DeepSeek V4 FlashDeepSeek | Chat | 1M tokens | $0.1526/M in · $0.3052/M out |
 Mistral SmallMistral | Chat | 128k tokens | $0.109/M in · $0.327/M out |
Why these models
Flash/mini/Haiku for ~90% of questions; Pro/Sonnet when retrieval is noisy or the answer is critical. Pair with the Embedding Models collection.
Use them with Geek Hub
One OpenAI-compatible base URL and API key. Swap any model id from this list without rewriting your client.
Get an API keyFAQ
- Which embedding with these?
- text-embedding-3-small by default. See Embedding Models. Do not mix dimensions in one index.
- Long context instead of RAG?
- Sometimes. If the corpus fits and input cost is fine, long-context can work. RAG scales better to GBs of docs.
- How do I reduce hallucinations?
- Strict prompt (“only use the context”), decent top-k, and an obedient model (Sonnet/GPT mini). Measure with groundedness evals.