RAMANUJAN

Intent-aware answer compression · WebGPU via Forge

Ramanujan is a portable neural network for intent-aware neural answer compression.

AI tools return long, repetitive output. Ramanujan reads it along with the questions the user has answered, picks the evidence that matters to them and writes one paragraph of 1–3 grounded sentences. When the user changes an answer, the summary changes too, and the tools don't run again.

AI tools

Search, retrieval and coding agents. They supply the information.

Harness

Runs the tools, asks the user questions and sends a structured request.

Ramanujan

Picks what matters for this user and says it in 1–3 sentences, citing the evidence.

Forge

Runs the encoder–decoder on WebGPU in the browser from one FZM checkpoint.

How it works with other AI

A scripted walkthrough. Answer the harness's questions in any order, or change an answer, and watch the evidence and the summary change. The tool outputs stay the same.

1User → AI agent

Which open-weight model should we use for our customer-support bot?

2AI tool outputs → harness

3Harness asks the user

4Ramanujan → harness

    Show the request and response JSON 
    

    

    Illustrative only: these outputs are scripted. Model names and facts are example tool output.
    

    


    

    QESC: Query-Guided Evidence Selection and Compression
    

    An encoder–decoder transformer with an evidence-selection head: 6 encoder and 4 decoder layers, hidden size 384, 6 heads, a 32K vocabulary, up to 2,048 input tokens and 96 output tokens.
    

    


Tool outputs
Selected QA pairs
Segmentation≤64 sentences + spans
Source encodercached on refinement
Intent encoderselected vs unselected
Contextual fusion
Evidence selectorrelevance = MLP(sentence, intent)
Redundancy scoringdrop near-duplicates
Top-k evidence3–8 sentences, provenance kept
Transformer decodercross-attention, greedy
1–3 sentences + evidence





    

    Source encoder
    Encodes every tool output and produces one representation per sentence, keeping each sentence's source id and byte span.
    

    Intent encoder
    Encodes the QA pairs the harness collected and keeps the selected ones apart from the rest. No hand-written prompt is needed.
    

    Evidence selection
    Scores each candidate sentence against the intent, lowers the score of sentences that repeat evidence already chosen, and keeps the top k.
    

    Decoder
    Generates one paragraph with cross-attention over the selected evidence and the intent, and keeps it to 1–3 sentences.
    

    Refinement
    A new QA selection reuses the cached source encoding. Only the intent, the evidence and the summary are recomputed.