🔧 Basic Syntax

Query the Space Race corpus with semantic search + LLM answering. Use the .venv312 interpreter — semantic search needs the local MiniLM model (torch is unavailable on Python 3.14/Intel Mac). Keyword mode works in any venv.

🎯 Keyword Mode (exact match, no API)

--keyword, -k — exact substring search over the corpus. Fast, deterministic, needs no embeddings.

--context, -c — turns of context before/after each match (default 3)

📌 Simple Examples

🎯 Filtering Options

--mission, -m — Filter by mission (e.g., as11, as13, as8)

--speaker, -s — Filter by speaker (e.g., cdm, cc, lmp)

🔍 Output Modes

--mode, -M — Output format: 'answer' or 'dialogue' (default: answer)

🔧 Advanced Options

--top-k — Number of turns to retrieve (default: 5)

--verbose, -v — Show similarity scores and reasoning

--debug — Show full system + user prompts

--log-level — Logging: DEBUG, INFO, WARNING (default: INFO)

📊 Complete Example

✅ Features

✓ Semantic search — local MiniLM embeddings (all-MiniLM-L6-v2, 384-dim), no API needed
✓ Keyword exact-match search over the full corpus
✓ LLM answer generation — Gemini (gemini-flash-latest)
✓ Context assembly (turns before/after for continuity)
✓ Similarity scoring (cosine, 0–1 scale)
✓ Mission + speaker filtering (indexed in Qdrant)
✓ Answer or dialogue (scene) output modes

🌐 Web UI

This page can run queries directly when the backend is up:

Then open http://localhost:8000. The server keeps the corpus and the embedding model in memory, so web queries skip the per-invocation startup cost the CLI pays.