Cuts 10-K research from hours to seconds — every answer cited back to the exact filing sentence.
Query → Classify (1 LLM call) → FAISS Retrieve → Generate (1 LLM call) → Cited Answer
2 API calls per query. The original pipeline used 7 (1 classify + 5 grade + 1 generate). I moved chunk grading into the generation prompt — the model self-filters irrelevant context as part of its reasoning. Same quality, 60% less latency.
| 8 companies indexed | Apple, Microsoft, Tesla, JPMorgan, Goldman Sachs, Amazon, NVIDIA, Alphabet |
| 2,450 document chunks | Parsed from real SEC EDGAR 10-K filings, section-aware |
| 2 API calls per query | Down from 7 — prompt engineering eliminated the grading step |
| ~2s average response | Groq inference is 10x faster than OpenAI for same model families |
| 14 LLM fallback models | User never sees a timeout — cascade handles rate limits silently |
| 18 Gemini extraction models | Source formatting always works, even on free-tier quotas |
Groq's free tier rate-limits aggressively. Instead of showing users a spinner or a "try again later" message, I built a 14-model cascade. If the best model (120B) is rate-limited, the request silently falls to the next best model. The user never waits.
Same approach for source extraction — 18 Gemini models chained so formatted source cards always render.
| Layer | Tech |
|---|---|
| Orchestration | LangGraph (deterministic state machine, not agents) |
| Vector search | FAISS + MiniLM-L6-v2 (384-dim embeddings) |
| LLM inference | Groq API (14-model cascade) |
| Source formatting | Google Gemini (18-model cascade) |
| Data pipeline | SEC EDGAR API + BeautifulSoup |
| Frontend | Streamlit |
git clone https://github.com/Vedag812/finsight-sec-rag.git
cd finsight-sec-rag
pip install -r requirements.txt
cp .env.example .env # add your Groq + Gemini keys
python scripts/ingest.py
streamlit run src/app/main.pyKeys needed: Groq (free) + Google AI Studio (free)
src/
├── pipeline/
│ ├── graph.py # LangGraph: Classify → Retrieve → Generate
│ ├── llm.py # Groq client + 14-model fallback
│ └── prompts.py # XML-delimited prompt templates
├── ingestion/
│ ├── sec_downloader.py # SEC EDGAR downloads
│ ├── document_parser.py# Section-aware 10-K parsing
│ └── chunker.py # Semantic chunking with metadata
├── vectorstore/
│ ├── embedder.py # SentenceTransformer embeddings
│ └── faiss_store.py # FAISS index operations
└── app/
├── main.py # Streamlit entry point
└── components.py # UI, CSS, Gemini source extraction
Built by Vedant Agarwal · MIT License