GitHub - ernes-toe/pulse-hermes: An AI research engine scored by upvotes, likes, and real money — not editors. Search Reddit. Hacker News. Polymarket prediction markets. YouTube. GitHub. The web. News. Score it all by what real people actually engage with. Rank it with Weighted Reciprocal Rank Fusion. Cluster the results thematically. Deliver a research briefing in seconds.

PULSE v4.0 — The Pulse of the Internet

⚠️ NOT MAINTAINED BY HUMANS. EVOLVED BY MACHINES.

PULSE is a multi-source social search engine scored by real engagement — not SEO, not editors, not algorithms designed to sell ads. It searches 18 platforms simultaneously and ranks by what real people actually engage with.

The human built the floor. The agents building the cathedral.

What Changed in v4.0

Six upgrades from a three-wave arXiv deep research session (~93 papers analyzed):

1. Query Router

Classifies queries into 5 types and routes to optimal sources automatically:

breaking_news → Reddit, HN, News, RSS, Bluesky, Lemmy
academic_deep → ArXiv, Semantic Scholar, OpenAlex, GitHub
prediction → Polymarket, Metaculus, Manifold, Reddit
technical_comparison → GitHub, ArXiv, HN, StackExchange, Dev.to
sentiment_pulse → Reddit, Bluesky, Lemmy, Hacker News

2. Adaptive Lookback

Dynamic time window based on topic activity — hot topics get fresher results:

Hot (>100 results/7d) → 7-day window
Active (>20 results/14d) → 14-day window
Default → 30-day window

3. Dedup Upgrade

Bigram pre-filter before cosine similarity. O(n log n) instead of O(n²). Adaptive threshold (0.90/0.85/0.80). Source-aware: higher engagement wins across duplicate sources.

4. Iterative Retrieval

Multi-round research with perspective gap analysis. Tracks 6 perspective categories (community, expert, academic, market, news, code). Stops when coverage hits 70% or rounds exhausted.

5. Trend Detection

Velocity spikes, source spread, keyword drift detection. --breaking mode: polls every 5 minutes, alerts on 3x velocity spikes.

6. Multi-Agent Research Crew

Four specialized agents: Collector → Analyzer → Specialist → Synthesizer. CoverageRubric scores 0-100. Automatic gap filling across rounds.

Scoring Weights (v4.0)

Signal	Weight	Change
Local Relevance	25%	↓ from 30%
Freshness	15%	↓ from 20%
Engagement	20%	↓ from 25%
Engagement Velocity	10%	—
Source Quality	10%	↓ from 15%
Retentive Value	10%	NEW
Cross-Source Confirmation	10%	NEW

An AI research engine scored by upvotes, likes, and real money — not editors.

Search Reddit. Hacker News. Polymarket. YouTube. GitHub. ArXiv. Lobsters. RSS. The web. News. Bluesky. Dev.to. Lemmy. OpenAlex. Semantic Scholar. StackExchange. Manifold. Metaculus. Tickertick. Bing News. SerpAPI News. All at once.

Score it all by what real people actually engage with. Rank it with Weighted Reciprocal Rank Fusion. Cluster the results thematically. Deliver a research briefing in seconds.

Google aggregates editors. PULSE searches people.

Sources

Source	What It Tells You	Auth
Reddit	The unfiltered take. Top comments with upvote counts.	No
Hacker News	The developer consensus. Points and comments.	No
Polymarket	Not opinions. Odds. Backed by real money.	No
YouTube	The 45-minute deep dive. Full transcripts searched.	No
ArXiv	Academic papers. ML/AI research. Peer-reviewed signal.	No
Lobsters	Curated tech links. Systems programming. Quality community.	No
RSS/Blogs	Technical blogs. Engineering insights. Expert opinions.	No
StackExchange	Q&A from Stack Overflow and the SE network. Expert answers.	No
Bluesky	Decentralized social. Growing alternative to X/Twitter.	No
Dev.to	Developer blog posts. Low-barrier technical content.	No
Lemmy	Decentralized Reddit alternative. Alternative community takes.	No
OpenAlex	Open academic graph. 250M+ scholarly works.	No
Semantic Scholar	Academic papers with citation data and recommendations.	No
Manifold	Play-money prediction markets. Community forecasts.	No
Metaculus	Forecasting community. Calibrated prediction track records.	No
Tickertick	Curated news aggregation. Topic-based news feeds.	No
GitHub	PR velocity, top repos by stars, issues, release notes.	Token
Web	Editorial coverage, blog comparisons.	Key
News	NewsAPI articles from major publications.	Key
Bing News	Microsoft Bing News search results.	Key
Serper News	Google News via Serper.dev API.	Key

18 sources. 16 work without any API keys.

Quick Start

git clone https://github.com/itsXactlY/pulse-hermes && cd pulse-hermes
bash install.sh

pulse "your topic" --yolo           # Autonomous run
pulse "bitcoin halving 2028" --depth deep --yolo
pulse --diagnose                     # Show available sources
pulse --setup                        # First-run wizard

No Install (Direct)

python3 scripts/pulse.py "your topic" --yolo

CLI Reference

pulse <topic> [options]

Options:
  --emit MODE       Output: compact (default), json, full, context, md
  --depth MODE      Research depth: quick, default, deep
  --sources LIST    Comma-separated sources
  --lookback N      Days to look back (default: 30, adaptive if auto)
  --save-dir DIR    Save report to directory
  --yolo            Skip human approval — run fully autonomous
  --crew            Multi-agent deep research (Collector→Analyzer→Specialist→Synthesizer)
  --iterative       Multi-round retrieval with perspective gap filling
  --max-rounds N    Max rounds for --crew/--iterative (default: 3)
  --breaking        Breaking-news monitor: poll every 5 min, alert on spikes
  --diagnose        Show environment and available sources
  --setup           Run first-run setup wizard
  --stats           Show cache and store statistics
  --history TOPIC   Show research history for a topic
  --trending        Show trending findings across topics
  --no-llm          Disable LLM planner (use heuristic)
  --no-cache        Disable cache (always fetch fresh)
  --no-store        Disable persistent store
  --no-progress     Disable progress display
  --debug           Enable debug logging

Configuration

Optional API keys go in ~/.config/pulse/.env:

# Web search (pick one):
BRAVE_API_KEY=***          # Free: 2000 queries/month
SERPER_API_KEY=***         # Google search via serper.dev
EXA_API_KEY=***            # Semantic search via exa.ai

# GitHub:
GITHUB_TOKEN=***           # Or use `gh auth login`

# News:
NEWSAPI_KEY=***            # Free: 100 requests/day

# LLM Planner (optional — Ollama is auto-detected):
OPENROUTER_API_KEY=***     # Cheapest cloud LLM
OPENAI_API_KEY=***         # OpenAI GPT-4o-mini

Or run pulse --setup for the interactive wizard.

How Scoring Works

Seven signals fused via Weighted Reciprocal Rank Fusion (RRF, k=60):

Signal	Weight	What It Measures
Local Relevance	25%	Token overlap with topic (bigram/trigram weighted)
Freshness	15%	Recency within the lookback window
Engagement	20%	Platform metrics (upvotes, points, volume, views, stars)
Engagement Velocity	10%	How fast engagement grows per day
Source Quality	10%	Baseline trust + self-learning weights
Retentive Value	10%	Did this source help with similar topics before?
Cross-Source Confirmation	10%	Same content from 3+ sources = higher trust

Four-pass deduplication: URL exact → Content hash → Bigram pre-filter → Cosine similarity.

Architecture

scripts/
  pulse.py                  # CLI entry point
  auto_commit.sh            # Auto commit + test + PR (for agents)
  hermes_bootstrap.sh       # Auto-discovery for new Hermes agents
  lib/
    schema.py               # Data models
    pipeline.py             # Orchestrator
    planner.py              # Heuristic planner
    llm_planner.py          # LLM planner (Ollama/OpenRouter/OpenAI)
    normalize.py            # Source normalizers
    score.py                # 7-signal scoring
    dedupe.py               # 4-pass dedup (URL, hash, bigram, cosine)
    fusion.py               # Weighted RRF
    cluster.py              # Cosine similarity clustering
    render.py               # Output rendering (5 modes)
    cache.py                # SQLite cache (24h TTL)
    store.py                # Persistent research store
    ui.py                   # Live progress display
    setup.py                # First-run setup wizard
    config.py               # Environment management
    query_router.py         # Query type classification + source routing
    adaptive_lookback.py    # Dynamic time windows
    iterative_retrieval.py  # Multi-round gap analysis
    trend_detector.py       # Velocity/spread/drift detection
    research_crew.py        # Multi-agent research pipeline
    relevance.py            # Token overlap + cosine similarity
    self_learn.py           # Self-learning source weights
    neural_memory.py        # Neural memory integration
    filter.py               # Result filtering
    raw_filter.py           # Raw data pre-filtering
    http.py                 # HTTP client utilities
    log.py                  # Structured logging
    dates.py                # Date/time helpers
    # Sources (18):
    reddit.py, hackernews.py, polymarket.py, youtube.py,
    github.py, web_search.py, news.py, arxiv.py, lobsters.py, rss.py,
    bluesky.py, devto.py, lemmy.py, openalex.py, sem_scholar.py,
    stackexchange.py, manifold.py, metaculus.py, tickertick.py,
    bing_news.py, serpapi_news.py

Output Modes

Mode	Description	Use Case
`compact`	Terminal-friendly ranked clusters	Quick research, CLI
`full`	All items by source, detailed	Auditing, saving
`json`	Machine-readable structured data	Programmatic use
`context`	Compact snippet for embedding	Agent integration
`md`	Markdown report	Documentation, sharing

Comparison with the Original

Feature	mvanhorn/last30days	PULSE v4.0
Sources	14+	18
Free sources (no keys)	3	16
Query Router	✗	✓ (5-route auto-classification)
Multi-Agent Crew	✗	✓ (Collector→Analyzer→Specialist→Synthesizer)
Iterative Retrieval	✗	✓ (gap analysis, early stop)
Trend Detection	✗	✓ (velocity, drift, breaking mode)
Adaptive Lookback	✗	✓ (7/14/30 day auto)
7-Signal Scoring	✗	✓
Cross-Source Confirmation	✗	✓
Retentive Relevance	✗	✓
Bigram-Accelerated Dedup	✗	✓ (O(n log n))
Neural Memory	✗	✓
Self-Learning	✗	✓
Python stdlib only	✓	✓
Lines of code	~15,000+	~11,000

Contributing

bash scripts/auto_commit.sh "feat: add Mastodon source"

Tests run. Commit happens. PR opens. Other agents review.

License

MIT — see LICENSE.

The human built the floor. The cathedral is yours.

Name		Name	Last commit message	Last commit date
Latest commit History 31 Commits
.github/workflows		.github/workflows
assets		assets
scripts		scripts
tests		tests
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
SKILL.md		SKILL.md
install.sh		install.sh
pyproject.toml		pyproject.toml

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

What Changed in v4.0

1. Query Router

2. Adaptive Lookback

3. Dedup Upgrade

4. Iterative Retrieval

5. Trend Detection

6. Multi-Agent Research Crew

Scoring Weights (v4.0)

Sources

Quick Start

No Install (Direct)

CLI Reference

Configuration

How Scoring Works

Architecture

Output Modes

Comparison with the Original

Contributing

See Also

License

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Folders and files

Latest commit

History

Repository files navigation

What Changed in v4.0

1. Query Router

2. Adaptive Lookback

3. Dedup Upgrade

4. Iterative Retrieval

5. Trend Detection

6. Multi-Agent Research Crew

Scoring Weights (v4.0)

Sources

Quick Start

No Install (Direct)

CLI Reference

Configuration

How Scoring Works

Architecture

Output Modes

Comparison with the Original

Contributing

See Also

License

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages