Retrieval, tools, or fine-tuning: how to decide for a voice agent

Imagine building a support voice agent from three files: an FAQ, an orders export, and customer conversations.

The FAQ contains facts the agent should retrieve. Order status needs a live lookup tool. The conversations might help with behavior, but first you need to test whether a prompt, the right facts, and working tools already solve the problem.

That decision is what I’m building vakforge for.

Its current CLI inspects your data and recommends a route. In a small synthetic shop example, it finds an FAQ, two tables, and 16 chat messages. It suggests retrieval and possible lookups—and says there isn’t enough evidence to fine-tune yet.

I’m planning to open-source the pre-alpha project tomorrow. It doesn’t train or serve a voice agent today; it helps answer the question that comes before training.

If you’ve built an agent from company data, which decision took longer than you expected?