1<h1 align="center">llm_aggregator</h1>
  2
  3<p align="center">
  4    <img src="assets/logo.svg" alt="LLM Aggregator logo" width=500>
  5    <br>
  6    <strong>A CLI tool to aggregate RSS feeds and summarise them with LLMs</strong>
  7</p>
  8
  9![Codeberg Release](https://img.shields.io/gitea/v/release/maxwelljensen/llm_aggregator?gitea_url=https%3A%2F%2Fcodeberg.org&style=for-the-badge)
 10![Codeberg License](assets/eupl-12-badge.svg)
 11
 12---
 13
 14## What is `llm_aggregator`?
 15
 16`llm_aggregator` is a command‑line utility that fetches articles from multiple
 17RSS feeds, filters and processes the content, and sends it to an LLM through
 18OpenAI-compatible API to generate a concise summary or analysis. It’s designed
 19for keeping up with news and articles from your favourite sources without
 20having to read dozens or hundreds of individual posts.
 21
 22## How do I use `llm_aggregator`?
 23
 24    llm_aggregator --feeds-file FEEDS-FILE --prompt PROMPT [OPTIONS]...
 25
 26By default, `llm_aggregator` reads a list of RSS feed URLs, fetches the
 27articles, filters them by date and keywords, and sends a summary request to
 28the LLM. The resulting output is printed to the terminal in your chosen format
 29(text, markdown, or JSON).
 30
 31### Basic Options
 32
 33    -f, --feeds-file FILE    Path to file containing RSS feed URLs (one per line)
 34        --stdin              Read a single RSS/Atom feed from stdin (combinable with -f)
 35    -p, --prompt PROMPT      User prompt for summarisation/analysis                 [required]
 36    --api-key KEY            API key (default: read from $LLM_AGGREGATOR_API_KEY)
 37    -m, --model MODEL        Model to use (default: deepseek-chat)
 38    --base-url URL           API base URL (default: https://api.deepseek.com)
 39    --max-tokens N           Maximum tokens in response (default: 4000)
 40    -P, --plain              Output only the raw LLM response without formatting or metadata
 41    -o, --output FORMAT      Output format: text, markdown, or json (default: text)
 42    --output-file FILE       Write output to FILE instead of STDOUT
 43    -t, --tui                Enable TUI interface with progress bar
 44    -D, --dry-run            Validate config, show article statistics, and exit without making LLM API calls
 45    -v, --verbose            Enable verbose logging
 46    -h, --help               Show this help message and exit
 47    --version                Show version information and exit
 48
 49### Filtering & Processing
 50
 51    -n, --max-articles-per-feed N  Maximum articles to fetch from each feed (default: 10)
 52    -d, --max-days-old N           Only include articles from the last N days (0 for all) (default: 7)
 53    --max-total-articles N         Maximum total articles to process (default: 20)
 54    -i, --include-keywords LIST    Comma-separated list of keywords to include (case‑insensitive)
 55    -e, --exclude-keywords LIST    Comma-separated list of keywords to exclude (case‑insensitive)
 56    --include-articles             Include original articles in JSON output
 57
 58### LLM Configuration
 59
 60    --temperature VALUE       Sampling temperature (0.0 to 1.0) (default: 0.7)
 61    --system-prompt TEXT      Custom system prompt for LLM
 62
 63### Examples
 64
 65```bash
 66# Basic usage: summarise tech news from a list of feeds
 67llm_aggregator -f feeds.txt -p "What are the latest AI-related trends in free software?"
 68
 69# Read RSS feed directly from stdin
 70curl -s https://example.com/feed.xml | llm_aggregator --stdin -p "Summarise"
 71
 72# Combine feeds file with stdin feed
 73llm_aggregator -f feeds.txt --stdin -p "Summarise all"
 74
 75# With TUI progress bar
 76llm_aggregator -f feeds.txt -p "Summarise tech news" -t
 77
 78# Output to a JSON file with included articles
 79llm_aggregator -f feeds.txt -p "Analyse AI developments" \
 80    -o json --output-file analysis.json --include-articles
 81
 82
 83# Filter by keywords (only include articles about Linux or open source)
 84llm_aggregator -f feeds.txt -p "Linux news" \
 85    -i linux,opensource -d 3
 86
 87# Use a custom model and higher token limit
 88llm_aggregator -f feeds.txt -p "Code analysis" \
 89    -m deepseek-reasoner --max-tokens 8000
 90
 91# Use a custom API endpoint (e.g., local Ollama)
 92llm_aggregator -f feeds.txt -p "Summarise news" \
 93    --base-url "http://localhost:11434/v1" -m llama3
 94
 95# Show version information
 96llm_aggregator --version
 97
 98# Show help message
 99llm_aggregator --help
100```
101
102## How does `llm_aggregator` work?
103
104`llm_aggregator` performs the following steps for each run:
105
1061. Parse command‑line arguments
1072. Read the feeds file: a plain text file containing one RSS/Atom feed URL per
108   line.
1093. **Fetch and parse feeds concurrently**: RSS, Atom, and JSON Feed formats are
110   supported. Feeds are fetched in parallel with rate limiting to maximise
111   throughput while avoiding server overload.
1124. **Extract article content**: for each feed entry, the tool extracts the
113   title, link, publication date, author, and description. If the feed provides
114   only a snippet, it can optionally fetch the full webpage using `goquery` to
115   extract the main content.
1165. **Filter and sort articles**: articles are filtered by age (configurable
117   with `--max-days-old`), optionally filtered by keywords (include/exclude),
118   and sorted by date, title, or source.
1196. Prepare the prompt with selected articles, formatted into a context
120   string that is sent to the LLM along with the user’s custom prompt.
1217. Call the OpenAI API via the `openai‑go` client.
1228. **Format and output the result**: the AI’s response is printed in the chosen
123   format (plain text, GitHub‑flavoured markdown, or JSON). If JSON output is
124   selected, the original articles can be included alongside the summary.
125
126![LLM Aggregator action in GIF](./assets/demo.gif)
127
128When the `--tui` flag is used, the entire process is wrapped in a `bubbletea`
129TUI that shows a colourful progress bar, live article counters, and elapsed
130time. The TUI renders Markdown content from the LLM with proper styling for
131headers, bold, italic, code blocks, and lists. The TUI supports keyboard
132navigation (j/k, arrows, space, b, g/G) and mouse wheel scrolling for
133browsing long summaries.
134
135Feeds are fetched concurrently for optimal performance, with rate limiting to
136avoid overwhelming feed servers.
137
138## Configuration
139
140`llm_aggregator` supports multiple configuration sources with the following precedence order (highest to lowest):
141
1421. **Command‑line arguments** – Override everything
1432. **Environment variables** – Start with `LLM_AGGREGATOR_` prefix
1443. **Configuration file** – `~/.config/llm_aggregator/config.toml`
1454. **Built‑in defaults**
146
147### Configuration file
148
149Create a TOML file at `~/.config/llm_aggregator/config.toml` with the following structure:
150
151```toml
152# Feed aggregation options
153max_articles_per_feed = 10
154max_days_old = 7
155max_total_articles = 20
156
157# Content filtering (comma-separated keywords)
158# include_keywords = "linux,opensource"
159# exclude_keywords = "windows,microsoft"
160
161# LLM API options
162# api_key = "your_api_key_here"  # Can also be set via LLM_AGGREGATOR_API_KEY env var
163# base_url = "https://api.deepseek.com"  # Optional custom API endpoint
164model = "deepseek-chat"
165max_tokens = 4000
166temperature = 0.7
167
168# System prompt for LLM API
169system_prompt = """You are an expert analyst and summariser.
170You analyse content from multiple sources and provide
171concise, insightful summaries based on user requests.
172Focus on key points, trends, and important information."""
173
174# Output options
175output = "text"  # Options: text, json, markdown
176# output_file = ""  # Optional output file path
177include_articles = false
178```
179
180### Environment variables
181
182All configuration options can also be set via environment variables with the `LLM_AGGREGATOR_` prefix:
183
184| Variable | Description |
185|----------|-------------|
186| `LLM_AGGREGATOR_API_KEY` | LLM API key |
187| `LLM_AGGREGATOR_BASE_URL` | API base URL (default: "https://api.deepseek.com") |
188| `LLM_AGGREGATOR_MODEL` | Model name (default: "deepseek-chat") |
189| `LLM_AGGREGATOR_MAX_TOKENS` | Maximum tokens in response (default: 4000) |
190| `LLM_AGGREGATOR_TEMPERATURE` | Sampling temperature (default: 0.7) |
191| `LLM_AGGREGATOR_SYSTEM_PROMPT` | Custom system prompt |
192| `LLM_AGGREGATOR_MAX_ARTICLES_PER_FEED` | Maximum articles per feed (default: 10) |
193| `LLM_AGGREGATOR_MAX_DAYS_OLD` | Maximum article age in days (default: 7) |
194| `LLM_AGGREGATOR_MAX_TOTAL_ARTICLES` | Maximum total articles (default: 20) |
195| `LLM_AGGREGATOR_INCLUDE_KEYWORDS` | Comma‑separated include keywords |
196| `LLM_AGGREGATOR_EXCLUDE_KEYWORDS` | Comma‑separated exclude keywords |
197| `LLM_AGGREGATOR_OUTPUT` | Output format (default: "text") |
198| `LLM_AGGREGATOR_OUTPUT_FILE` | Output file path |
199| `LLM_AGGREGATOR_INCLUDE_ARTICLES` | Include articles in JSON output (true/false) |
200| `LLM_AGGREGATOR_PLAIN` | Plain output without metadata (true/false) |
201| `LLM_AGGREGATOR_STDIN` | Read RSS/Atom feed from stdin (true/false) |
202
203The API key can be provided via `--api‑key`, `LLM_AGGREGATOR_API_KEY` environment variable, or in the configuration file.
204
205## Example feeds file
206
207Create a file named `feeds.txt` with your favourite RSS feeds, one per line.
208For example:
209
210    https://news.ycombinator.com/rss
211    https://lwn.net/headlines/newrss
212    https://opensource.com/feed
213    https://www.phoronix.com/rss.php
214
215Then run:
216
217    llm_aggregator -f feeds.txt -p "Summarise the top tech stories"
218
219## Dependencies
220
221`llm_aggregator` is written in Go and uses the following libraries:
222
223| Library | Description |
224|---------|-------------|
225| [`gofeed`](https://github.com/mmcdole/gofeed) | Robust RSS/Atom/JSON feed parser |
226| [`openai‑go`](https://github.com/openai/openai-go) | Official OpenAI API library for Go |
227| [`bubbletea`](https://github.com/charmbracelet/bubbletea) | TUI framework for terminal applications |
228| [`lipgloss`](https://github.com/charmbracelet/lipgloss) | Library for styling terminal output |
229| [`glamour`](https://github.com/charmbracelet/glamour) | Markdown rendering for terminal (used in TUI mode) |
230| [`go‑arg`](https://github.com/alexflint/go-arg) | Struct‑based argument parsing |
231| [`tiktoken-go`](https://github.com/pkoukk/tiktoken-go) | OpenAI's tiktoken BPE tokeniser |
232| [`viper`](https://github.com/spf13/viper) | Configuration management |
233| [`goquery`](https://github.com/PuerkitoBio/goquery) | jQuery‑like HTML scraping |
234
235## How do I build `llm_aggregator`?
236
237`llm_aggregator` can be built with a standard Go toolchain:
238
239    go build ./cmd/llm_aggregator.go
240
241For information about the project's test suite, see
242[docs/TESTING.md](docs/TESTING.md).
243
244## Licence
245
246This project is licensed under [European Union Public Licence
2471.2](https://joinup.ec.europa.eu/collection/eupl/eupl-text-eupl-12).