π€ Multilingual RAG + Ensemble Sentiment + Economic Forecast
FinBERT (40%) + XLM-RoBERTa (30%) + Loughran-McDonald (30%)
Upload WEO reports β Build Index β Chat β Forecast
Upload PDF / DOCX / TXT / CSV
π‘ Name files like
WEO_2020.pdfβ year in filename required! β Year must show β not β οΈ
No index yet.
Ask anythingβ¦ / Ψ§Ψ³Ψ£Ω Ψ£Ω Ψ΄ΩΨ‘β¦
π― Similarity Threshold
0.05 0.5
π§ Tech Stack
| Component | Model / Tool |
|---|---|
| π¦ Sentiment 1 | FinBERT β ProsusAI (40%) |
| π Sentiment 2 | XLM-RoBERTa β Cardiff NLP (30%) |
| π Sentiment 3 | Loughran-McDonald Lexicon (30%) |
| β‘ Ensemble | Weighted 40/30/30 + MinMax[-0.3,+0.3] |
| π Embeddings | paraphrase-multilingual-MiniLM-L12-v2 |
| π Reranker | cross-encoder/ms-marco-MiniLM-L-6-v2 |
| π€ LLM | Llama-3.3-70B via Groq |
| ποΈ STT | Whisper small |
| π TTS | gTTS |
| ποΈ Vector DB | FAISS IndexFlatIP |
| π Keyword | BM25 Hybrid |
| π Forecast | ARIMA(1,1,1) vs SARIMAX+Ensemble |
| π Data | World Bank API |
| πΎ Persistence | Pickle + FAISS |
π Country Codes
| π©πΏ Algeria | DZ | πΊπΈ USA | US |
| π«π· France | FR | π²π¦ Morocco | MA |
| πΉπ³ Tunisia | TN | πͺπ¬ Egypt | EG |
π ARIMA vs SARIMAX + Ensemble Sentiment
Steps:
- Upload
WEO_2019.pdf,WEO_2020.pdf,WEO_2021.pdfβ¦ - Build Index β verify Year = β
- Select country + variable + year range
- Run Forecast β compare RMSE
π¦ FinBERT(40%) + π XLM(30%) + π Lexicon(30%) Normalized to [-0.3, +0.3] for smooth forecasting
π― Target Variable
2000 2020
2010 2024