LeoAi_
1 MCP server offers this tool
What it usually takes
compare_mode– in 1 of 1enabled_sources– in 1 of 1history– in 1 of 1message– in 1 of 1retrieval_k– in 1 of 1strict_mode– in 1 of 1tone– in 1 of 1use_multi_query– in 1 of 1use_reranker– in 1 of 1
No server marks it as only reading: treat it as able to change data.
From each publisher's own tool list. We never call a tool.
Servers
| Server | Inputs | Read-only? |
|---|---|---|
LeoAi Lite co.huggingface.rodmanai/leoai-lite Fetch a reply: route → retrieve (optionally multi-query) → rerank → prompt → yield. Yields twice — first a "_Thinking..._" placeholder so the UI shows a pending turn while inference runs, then the full answer. flan-t5 via HuggingFacePipeline doesn't expose token-level streaming, so this is a two-yield interaction rather than a true stream. Modes that change behavior: - `use_multi_query`: rephrase the question 3 ways (one extra LLM call) and union retrievals from each — better recall. - `compare_mode`: force balanced per-corpus retrieval and switch to a side-by-side comparison template that splits Leo's chunks and the visitor's chunks into separate prompt sections. No-ops to regular behavior when the visitor has no uploads. - `strict_mode`: refuse to answer outside the retrieved context. The cross-encoder reranker (when enabled) reorders chunks and supplies the scores for the confidence label. Per-stage wall-clock timings are surfaced in the sources panel header. Each completed turn is | compare_mode,enabled_sources,history,message,retrieval_k,strict_mode,tone,use_multi_query,use_reranker | Not stated |