The feed and the knowledge graph

How does the feed decide what to show me? Cosine similarity against your interests profile until you have clicked anything, then a blend with two click-trained terms that phases in as evidence accumulates — plus click provenance, so only trustworthy feedback trains the ranker.

Feed ranking

The original core, still here. An agent orchestrator for personalized feed recommendations coordinating three layers with distinct reliability contracts: deterministic ingest (fetch → dedup → embed), a per-user learnable ranking core (profile embedding + click-trained classifier), and a LangGraph explain agent that says why items rank where they do — lazily, cached, never blocking the feed. The LLM is a swappable OpenAI-compatible backend, not the point.

It runs two ways, and they share one database:

1. Standalone — a web UI at http://127.0.0.1:8899 with ✓/✗ feedback buttons that retrain the ranker on every click.
2. Alongside hermes-agent — as an MCP stdio server exposing feed.* tools as native tool calls (see the agents guide for the full table), plus an optional attestation-setup skill for setup automation and fallback, and the attestation-feed skill for how an agent should behave with the feed.

How ranking works

• 0 clicks: cosine similarity between item embeddings (embeddinggemma, 256-dim) and your interests profile text.
• With clicks, two click-driven terms join the blend (weight w = n_clicks / (n_clicks + 5), averaged when both are active):
• a per-user logistic regression over item embeddings (guard: this term only participates once your clicks contain both classes);
• a feature-preference term — Laplace-smoothed like/dislike ratios per LLM-extracted topic tag, content type, and source feed (see uv run attest tag). This term works from click one, including for users who have only ever downvoted: two ✗ on items sharing a tag demote every item carrying that tag on the next render.
• Visible movement by click 3-4; tag-level demotion is visible immediately.

Click provenance

Every recorded click stores its provenance, and provenance decides what a row may be used for:

source	what it is
ui	you pressed a button on the web page
agent	an MCP feed.rate call, usually the agent reading your reply
implicit	you asked why an item ranked; engagement, counted as a weak positive
simulated	a local model reacting to the text as the persona would
bootstrap	synthetic persona seeding

bootstrap labels are a linear threshold on the same embedding the ranker's classifier consumes, so scoring against them is a tautology and evaluate_user excludes them. The other four are trainable, and feed.persona_status breaks the counts down by source so you can see how much of a persona's history is yours.

Explicit feedback is scarce by nature — this project's own database held 68 web clicks and 2 agent clicks across 5,167 items before the two synthetic channels were added, every one of them positive, which is why the click classifier had never fired for a real account.

The knowledge graph that sits beside the feed — concepts, centrality, communities — is derived from the same tagging pass; see the agents guide for how it is built and the full kg.* tool table.