AICryptoPulse

AICryptoPulse is an advanced Retrieval-Augmented Generation (RAG) system designed to curate and analyze daily crypto news. It powers the Telegram bot @agent_cryptopulse_bot, providing insightful updates directly to your chat.

---

🚀 Project Structure

• /bot: Telegram Bot User Interface (UI).
• /data: Airflow infrastructure for collecting and processing news feeds.
• /notebooks: Jupyter notebooks for research and experiments.
• /service: Core logic implementing the RAG pipeline and API application.

---

🛠 How to Run?

1. Set up the Airflow module (located in /data):

   • Refer to the official Airflow documentation for installation.
   • Run a PostgreSQL database to store feed data.
   • Set up S3-like bucket to store FAISS indexes.
   • Configure settings in /data/configs/.
   • Enable all DAGs in the Airflow interface.

2. Configure environment variables:

   • Use .env.example as a template to create your .env file.

3. Run the Service:

   • Use Docker Compose to deploy the system:

     docker-compose up -d

   • Alternatively, use the Makefile:

     make all

---

🌐 Data Pipeline

• Data is collecting from the open APIs (feeds, Twitter APIs, Telegram API)
• ETLs are running on Airflow and store all data in PostgreSQL
• FAISS index (both short-term and long-term) are updated each day on Airflow

Done:

• Coindesk
• DLNews
• Twitter big crypto accounts
• DeFillamaFeed

In progress:

• Tree Feed
• Custom Twitter accounts

To Do:

• Bloomberg
• Cointelegraph
• Classic Financial news portals

---

Benchmarks and metrics

• CryptoQA (HemaChandrao/crypto_QA) - synthetic QnA dataset with GPT answers (215 rows);
• Filtered crypto-2024 (sites.google.com/view/cryptoqa-2024/datasets) – dataset with answers to crypto questions from reddit and twitter from the Indian Institute of Technology (239 rows). Filtering was performed by filtering relevant items using the Qwen2.5-14b model.

HemaChandrao/crypto_QA

MTEB models

model_id	mAP	MRR
HIT-TMG/KaLM-embedding-multilingual-mini-instr...	0.818385	0.817797
jinaai/jina-embeddings-v3	0.807477	0.805951
Alibaba-NLP/gte-large-en-v1.5	0.779034	0.777258
WherelsAl/UAE-Large-V1	0.7547	0.752776
jxm/cde-small-v1	0.240464	0.224688

Base models

model_id	mAP	MRR
all-mpnet-base-v2	0.798267	0.796895
multi-qa-mpnet-base-dot-v1	0.778915	0.777039
all-distilroberta-v1	0.733526	0.730228
all-MiniLM-L12-v2	0.725623	0.722476
multi-qa-MiniLM-L6-cos-v1	0.716872	0.714196
multi-qa-distilbert-cos-v1	0.713394	0.710679
all-MiniLM-L6-v2	0.712731	0.709069
paraphrase-multilingual-mpnet-base-v2	0.610216	0.601813
paraphrase-albert-small-v2	0.607011	0.601449
paraphrase-multilingual-MiniLM-L12-v2	0.594264	0.585709
distiluse-base-multilingual-cased-v2	0.582778	0.575996
distiluse-base-multilingual-cased-v1	0.571691	0.563731
paraphrase-MiniLM-L3-v2	0.551712	0.543413

Filtered Cryptoqa-2024

MTEB models

model_id	mAP	MRR
Alibaba-NLP/gte-large-en-v1.5	0.631214	0.623856
HIT-TMG/KaLM-embedding-multilingual-mini-instr...	0.608928	0.602966
jinaai/jina-embeddings-v3	0.608519	0.601709
WherelsAl/UAE-Large-V1	0.554994	0.547885
jxm/cde-small-v1	0.155341	0.136702

Base models

model_id	mAP	MRR
all-mpnet-base-v2	0.575693	0.566497
all-distilroberta-v1	0.523694	0.512484
multi-qa-mpnet-base-dot-v1	0.515863	0.505068
all-MiniLM-L12-v2	0.509307	0.499188
all-MiniLM-L6-v2	0.469071	0.458595
multi-qa-distilbert-cos-v1	0.466012	0.453911
multi-qa-MiniLM-L6-cos-v1	0.434840	0.420884
distiluse-base-multilingual-cased-v1	0.305668	0.291416
paraphrase-multilingual-mpnet-base-v2	0.303580	0.287447
distiluse-base-multilingual-cased-v2	0.300423	0.283640
paraphrase-multilingual-MiniLM-L12-v2	0.274722	0.258638

---

RAG

Retriever - all-MiniLM-L6-v2.
Decoder - gpt-3.5-turbo.

Also our solution contains caching of model responses, for more reasonable spending of financial resources, chat history for the user.

---

📐 System Design

---

📏 Code Guidelines

• Separate Codebases: Clearly distinguish research code from production code.
• Lint Before Committing: Run linters using:

  make lint