Model guide

Connect a model, and FULI gains a skill.

FULI works with no model at all. A model is optional: it lets FULI find by meaning, sort projects and tidy memory.

Run a local model through Ollama, or use an online one. You approve which projects it may see.

Two ways to run FULI

Keyword ranking, plus the AI you already chat with.

Free, nothing to install.

Adds search by meaning, and the built-in Agents judge on their own.

A small model is enough for most uses.

Turns each knowledge item and each question into a vector, to find by meaning.

Enough: A small embedding model with Chinese support, 0.3B–0.6B.
Examples: Through Ollama: qwen3-embedding:0.6b (639 MB), embeddinggemma-2:270m (304 MB). Online embedding APIs work too.
Measured: Synthetic held-out set: 40 prompts, 58 items, top 3.

qwen3-embedding:0.6b32/32 found · 1 false hit in 8 · 16 ms per query (desktop CPU)

embeddinggemma-2:270m28/32 found · 1 false hit · 18 ms

Keyword ranking alone31/32 found on the same set
Tokens: No tokens with a local model.

Suggests which project a conversation belongs to.

Reads a project's recent sessions and proposes what to remember or retire. You confirm.

Enough

A strong chat model with a long context.

Measured

One full run over 21 real sessions of one project: 3 model calls, 143,292 tokens.

The daily run reads only what is new since the last run, so it costs much less.

Reviews knowledge before it is kept.

Enough

A strong chat model.

Measured

A local qwen3.5:4b agreed with the labels on 60.4% of 48 held-out cases, below our 85% bar.

So we recommend a strong online model, or the current session.

Matches work to board items.

Suggests one next step from an Agent's score card.

Synthetic data only, one project, small sets. The numbers are a guide, not a promise.