rss-llm-digest

An autonomous AI agent pipeline that aggregates RSS feeds, summarizes them with an LLM, and delivers formatted digests to Telegram or Email — orchestrated with LangGraph and LangChain.

---

AI Architecture

The pipeline is a multi-node stateful graph (LangGraph), not a single prompt call. Each stage is a specialized async node connected by a state machine with conditional routing.

START
  │
  ▼
summarizer_node   ← ReAct Agent (LangChain) + rss_feed tool + search_rss_history tool + DeepSeek LLM
  │
  ▼
integration_router  ← Conditional edge: routes by CHANNEL_TO_SEND
  │
  ├── telegram_node  ← RAG formatting + LLM reformat + aiogram delivery
  │
  └── email_node     ← MIME message + Gmail SMTP SSL

---

AI Patterns

ReAct Agent with Tool Calling — The summarizer node runs a LangChain agent equipped with two custom tools: rss_feed (fetches and stores RSS entries) and search_rss_history (searches past entries by semantic similarity). The agent autonomously decides what to read and what context to retrieve before summarizing.

Graph-Based Orchestration (LangGraph) — Nodes share typed state via a Pydantic State model. Conditional edges handle runtime routing between delivery channels. Adding a new channel means adding one node and one branch.

Retrieval-Augmented Generation (RAG) — Two separate Chroma collections running as a persistent external service:

• telegram_rules — Telegram Bot API HTML formatting spec, loaded once on first startup
• news_collection — all fetched RSS entries, accumulated across runs; used by the agent to find related articles and enrich summaries with historical links

Prompt Engineering — The LLM behavior is defined in a separate system_prompt.yaml with a strict output schema (TITLE, SUMMARY, WHY_IT_MATTERS, SOURCE, TARGET_AUDIENCE). Decoupled from code for independent versioning.

Scheduled Execution — The pipeline runs on a daily cron schedule via APScheduler (AsyncIOScheduler + CronTrigger). Send time is configured via SCHEDULE_TIME env var. The process runs continuously and waits for the next trigger.

Fully Async Pipeline — RSS fetching, LLM calls, and message delivery all run with asyncio and LangChain's async APIs.

---

Tech Stack

Layer	Technology
LLM	DeepSeek (deepseek-chat)
Agent Framework	LangChain — tool calling agent + AgentExecutor
Workflow Orchestration	LangGraph — StateGraph, conditional edges
Embeddings	HuggingFace all-MiniLM-L6-v2
Vector Store	Chroma (separate container, two persistent collections)
RSS Parsing	feedparser
Telegram Bot	aiogram 3.x (async)
Config	Pydantic BaseSettings
Logging	loguru
Tracing	LangSmith
Scheduling	APScheduler (AsyncIOScheduler + CronTrigger)
Containerization	Docker + Docker Compose + Poetry

---

Running

1. Configure environment

Create a .env file in the project root:

DEEPSEEK_API_KEY=...
LANGSMITH_API_KEY=...      # optional, for LangSmith tracing

CHANNEL_TO_SEND=TELEGRAM   # TELEGRAM | EMAIL
SCHEDULE_TIME=08:00        # daily digest time (HH:MM, system timezone)

# Telegram
TELEGRAM_BOT_TOKEN=...
TELEGRAM_CHAT_ID=...

# Email
EMAIL_FROM=...
EMAIL_TO=...
EMAIL_PASSWORD=...         # Gmail app password

2. Docker Compose (recommended)

docker-compose up --build

3. Local development

Run ChromaDB in Docker, app locally:

# Start ChromaDB
docker-compose up chroma

# Run the app
poetry install
poetry run python agent.py