Chat with Leo
Ask anything about Leo and the LeoAi-Lite project. Optionally upload your own documents — Leo will read them and answer questions about both. Tip: press ? for keyboard shortcuts.
Loading model (google/flan-t5-large)... first visit can take ~30-60s.
Reranks retrieved chunks by query relevance using cross-encoder/ms-marco-MiniLM-L-6-v2. Powers the confidence label.
Asks the LLM to rephrase your question 3 ways and unions retrievals from each. Adds one extra LLM call per turn.
After each answer, asks the LLM for 3 follow-up questions and shows them under the chat. Adds one LLM call per turn.
Forces balanced retrieval and uses a comparison prompt that contrasts Leo's background with your uploads. Only takes effect when at least one user source is active.
Try one of these:
Tell Leo who's asking and pick a register. Both shape how Leo replies but don't get treated as factual material to answer from.
Upload PDFs / DOCX / text files, paste text, or paste URLs (one per line — YouTube links are supported and pull the transcript). Each source becomes its own toggleable entry in the Active sources list below.
When on, the URLs box is ignored and no outbound HTTP happens. Files and pasted text still work.
Paste a job description. The bot extracts the top 5 requirements, retrieves matching evidence from Leo's corpus, scores each match with the cross-encoder, and returns a per-requirement table plus an overall fit verdict. One LLM call up front for extraction; rest is retrieval + reranker scoring.
Aggregated view of the question log: route distribution, confidence distribution, and questions per day. Refresh after each turn to watch numbers move. The flagged-answers list shows turns where the reranker confidence was Low — Leo's review pile.
Runs a small fixed Q&A set against Leo's corpus and reports whether the expected keywords were retrieved. Doesn't call the LLM — measures retrieval quality only. Re-run after changing k or toggling the reranker to see the impact.
Last 20 questions asked across all sessions on this Space instance, plus a top-N word-frequency view of what visitors are asking about. Stored as JSONL in /tmp — lost when the Space restarts.
Iframe embed — drop this into any page to render the chatbot. The Space URL is fixed via SPACE_EMBED_URL in app.py.
MCP server — this Space exposes Leo's ask_leo(question) function as an MCP tool. Add it to Claude Desktop / Cursor / Windsurf with this config (the exact endpoint may vary across Gradio + Spaces versions; check the Space's API docs at https://rodmanai-leoai-lite.hf.space/?view=api):
Want to talk to the real Leo? Visit rodman.ai