GitHub ↗
HOPPERPARSECHUNK800 chars120 overlapEMBEDSTAMPVECTOR VAULTBM25 CABINETEMBED QSEMKWFUSEPROMPTLLM

Send your document down the RAG factory line.

Every machine is a real pipeline step running on real data. Load files, ask a question, then tap any machine to open it.

READY · load files or ask a question

library
ANSWER TRAY
Answers land here, with numbered sources you can click.
no runs yet

Below the floor

Every station, explained.

  1. 01

    Parse

    Your file is turned into plain text. PDFs only store letters and where they sit on the page, so the reader rebuilds sentences in reading order.

    —
  2. 02

    Chunk

    The text is cut into ~800-character chunks, preferring paragraph and sentence breaks. Each chunk repeats the last 120 characters of the one before, so an idea that straddles a cut survives whole somewhere.

    —
  3. 03

    Embed

    An embedding model turns every chunk into 1,024 numbers: a point in "meaning space". Chunks about similar things land near each other.

    —
  4. 04

    Vector store

    The vectors are saved in Pinecone, a vector database, under a private namespace. The galaxy shows them squashed from 1,024 dimensions to 3 with PCA.

    —
  5. 05

    Keyword index

    In parallel every word is counted into a keyword index (BM25), the classic search-engine method. Words that are rare in the library but frequent in a chunk count the most.

    —
  6. 06

    Ask

    Your question enters the pipeline. The search mode decides which retrievers run: keyword, semantic, or both fused together (hybrid).

    —
  7. 07

    Embed question

    The question is embedded with the same model, so it becomes a point in the same space as the chunks: the amber probe in the galaxy.

    —
  8. 08

    Keyword search

    BM25 scores chunks by the words in your question. A rare word appearing often in a chunk scores high, which is why exact names, codes and jargon shine here.

    —
  9. 09

    Semantic search

    Vector search returns the 20 chunks whose vectors point the same way as the question (cosine similarity), even if they use different words.

    —
  10. 10

    Fuse (RRF)

    Reciprocal Rank Fusion merges both lists. Each chunk earns 1 / (60 + rank) from every list it appears in, so chunks both retrievers agree on rise to the top.

    —
  11. 11

    Rerank

    A cross-encoder reads the question and each candidate together and rescores it: slower than comparing vectors, but much sharper. The best 5 move on.

    —
  12. 12

    Prompt

    The winning chunks are numbered and packed into the prompt with instructions and your question, inside the model's token budget.

    —
  13. 13

    Generate

    The language model writes the answer one token at a time and cites the numbered chunks as [n]. Click a citation to see exactly where it came from.

    —
  14. 14

    Evaluate

    A second, smaller model grades the run: is each claim backed by the chunk it cites, does the answer address the question, were the retrieved chunks relevant?

    —
  15. 15

    Trace tape printer

    Every run is traced like a production service: how long each step took, how many tokens went in and out, and what it cost.

    —
  16. 16

    QA lab: which search wins?

    The same golden questions run through four retrieval set-ups. Recall@5: did the right chunk make the top 5? MRR: how high was the first right chunk? nDCG@10: how good is the whole top 10?

    —