032026

PaperForge AI

RESEARCH → RETRIEVAL → CODE

PythonRAGLLMsVector DBFAISS

RESEARCH PAPER

methodology, §3.2

RAG PIPELINE

retrieval.py

def retrieve(query, k=5):
    vec = embed(query)
    return faiss_index.search(
        vec, k
    )

chunk → embed → retrieve → generate

methodology_parser.py

class MethodParser:
    def to_python(self, eq):
        return compile_fn(eq)

(EVERYTHING HERE IS DRAGGABLE — GO AHEAD)

THE IDEA

Academic papers describe methods in prose and formulas that stay locked on the page. PaperForge AI reads a paper the way a researcher would — retrieve the relevant section, understand the method, then re-express it as code you can actually run.

THE PIPELINE

RESEARCH PAPER
DOCUMENT PROCESSING
SEMANTIC CHUNKINGrecursive character splitting
VECTOR SEARCHFAISS
LLM
METHODOLOGY
PYTHON CODE

THE SYSTEM

Documents are split with semantic chunking and recursive character splitting, tuned so a chunk stays a coherent unit of meaning rather than an arbitrary slice of text.
Chunks are embedded and indexed in a FAISS vector store for fast semantic retrieval.
A research-to-code parser takes the retrieved methodology and converts unstructured formulas into compilable Python functions, guided by strict prompts and output schemas.

THE PROCESS

Retrieval quality was evaluated with Recall@K to keep the pipeline honest about what it actually finds.
Prompt and schema design went through several passes to keep LLM output strictly structured — code that compiles, not code that merely reads well.

THE RESULT

A working pipeline: paper in, retrieved methodology and runnable Python out — built as a personal exploration of RAG systems applied to a genuinely hard retrieval problem.

papers → code