SEC 10-K RAG Pipeline

Summarize SEC 10-K filings with one function call.

Usage

from facade import summarize

summary = summarize(
    ticker="AAPL",
    item="1a",
    query="What are the main risk factors?",
    llm_api_key="your-mistral-api-key",
)
print(summary)

With Reranking

summary = summarize(
    ticker="AAPL",
    item="7",
    query="What drove revenue growth?",
    llm_api_key="your-mistral-key",
    reranker_api_key="your-zeroentropy-key",
)

Parameters

Item Sections

Installation

pip install -r requirements.txt

Configuration

Settings in config/settings.json:

{
  "chunking": {
    "chunk_size_tokens": 512,
    "chunk_overlap_tokens": 50,
    "min_chunk_chars": 50,
    "max_chunk_chars": 2200
  },
  "llm": {
    "model": "mistral-small-latest",
    "temperature": 0.3,
    "max_tokens": 1024
  },
  "retrieval": {
    "embedding_model": "BAAI/bge-small-en-v1.5",
    "vector_dim": 384,
    "top_k": 10,
    "min_score": 0.0,
    "rerank_top_k": 20
  }
}

Architecture

summarize(ticker, item, query, llm_api_key, reranker_api_key)
│
├─ 1. ingestion  → Download 10-K via datamule
├─ 2. chunking  → Split item section into chunks
├─ 3. storage        → Store chunks with embeddings
├─ 4. retrieval         → Vector search(sentence-transformers) and Optional reranking (Zero Entropy API)
└─ 5. inference         → Generate summary (Mistral API)