Agent settings

Eloquent gives you fine-grained control over how your agent uses the knowledge base.

Temperature

Think of temperature as the “creativity dial” of your agent.

• At 0, the agent decides itself whether to stick close to the source material or be a bit more free.
• At 0.1, the agent is extremely strict: it copies facts exactly as they are written in the knowledge base.
• At 1.0, the agent is highly creative: great for brainstorming, rewriting, or ideation.

In short: Lower values = safer, fact-based answers. Higher values = more flexible, creative output.

Similarity Threshold

This setting decides how close a piece of knowledge has to be to the user’s question before the agent uses it.

• A lower threshold means the agent accepts looser matches (faster answers, but sometimes irrelevant).
• A higher threshold means the agent only accepts very close matches (more precise, but might answer “I don’t know” more often).

In short: Use a higher threshold if accuracy is more important than always giving an answer.

Reranking

Most search systems grab results quickly, but don’t always understand intent. A reranker takes the first set of results and re-checks them with a smarter AI model. Instead of just matching keywords or vector similarity, it looks deeper at the relationship between the question and the content.

That means:

• More accurate answers for complex queries.
• Less “near misses” where the result looks close but isn’t really relevant.
• Especially useful when you have large or messy datasets.
• Reranking makes the agent a bit slower.

In short: Reranking makes sure the best results rise to the top, not just the first ones.

Number of Embeddings

Embeddings are the small chunks of knowledge your agent looks at before answering.

• More embeddings = more context, better coverage, but higher cost.
• Fewer embeddings = cheaper and faster, but the agent might miss some info.
• The sweet spot for most use cases is around 20 embeddings.

In short: Start with 20 and adjust based on cost vs. coverage.