Semantic vs LLM-Indexed Search Demo

Choose .txt file… No file chosen

Model Settings

Embedding model (all arms — each model keeps its own cached index)

all-MiniLM-L6-v223 MB · Fast, decent quality (default) all-MiniLM-L12-v233 MB · Slightly better quality all-mpnet-base-v286 MB · Higher quality, slower bge-small-en-v1.533 MB · Strong retrieval performance bge-base-en-v1.5109 MB · Best retrieval, slowest

LLM model (Arms B & C — each model keeps its own cached index)

SmolLM2-360M-Instruct360 MB · Fastest (default) SmolLM2-1.7B-Instruct1.7 GB · Better quality, ~5× slower Qwen2.5-0.5B-Instruct494 MB · Good balance of speed & quality Qwen2.5-1.5B-Instruct1.5 GB · High quality, slow

Transcript server (YouTube URL tab only — run server/main.py on your server)

Server URL

LLM Prompt Settings

Arms B and C run sequentially, not independently: Arm B summarizes each chunk first, then Arm C generates questions from Arm B's summaries (starting once a few are ready). Edit either prompt below before indexing to change what gets embedded for that arm and improve its search results — e.g. ask for more specific summaries, or questions phrased closer to how you actually search.

Arm B

Arm C

Cached Indexes

Arm A: Semantic

Query embedded and ranked by cosine similarity against raw chunk embeddings.

Arm B: Summary + RRF

Each chunk summarized in 2-3 sentences. RRF fuses raw chunk similarity and summary similarity.

Arm C: Questions + RRF

2-3 questions generated from each Arm B summary. RRF fuses raw chunk similarity and question similarity.