Earshot

Documentation

Open app

How embeddings work

A plain-language guide to how Earshot uses meaning-based matching — and what that means for your focus text, Knowledge Base, and subreddit picks.

The short version

Embeddings are how Earshot understands meaning, not just exact words.

When you write focus text, upload a Knowledge Base document, or click Suggest subreddits, Earshot converts that text into a kind of numerical "fingerprint" that captures what it's about. It does the same for Reddit posts and for chunks of your documents. Then it compares fingerprints to see what's related.

You never configure embeddings directly. But the text you write — focus descriptions, KB files, subreddit choices — is the input. Better inputs mean better matches.

A simple analogy

Imagine a library where books aren't shelved alphabetically, but by topic.

Books about "deployment pipelines" sit near books about "CI/CD" and "release automation" — even if they don't share the same words on the cover. Books about "celebrity gossip" sit somewhere else entirely.

Embeddings are Earshot's way of placing text on those shelves. Similar meaning → nearby. Different meaning → far apart.

That's why a post saying "our deploys keep breaking on Fridays" can match focus text about "CI/CD reliability" even when neither text uses the exact same phrases.

Keywords first, meaning second

Earshot uses a two-step filter on each Listener run. You'll see both stages on the Listener health funnel:

  1. Keyword filter — fast, literal matching. A post must hit your primary keywords (and respect negative terms) before it moves on.
  2. Contextual filter — embedding-based. Earshot compares the post's meaning to your Listener's Broad topics to focus on and decides if the conversation is actually on-topic.

Think of keywords as the bouncer at the door ("does this mention something we care about?") and embeddings as the host inside ("is this conversation actually what we're looking for?").

If you see lots of items passing keywords but failing the contextual filter, your focus text may be too narrow — or strictness may be too tight. See Listeners for tuning tips.

Where embeddings show up for you

1. Broad topics to focus on (contextual matching)

When you fill in Broad topics to focus on on a Listener, Earshot embeds that description and compares it to each candidate post or comment.

What this means for you:

  • Write focus text that describes the kind of conversation you want — problems, questions, comparisons — not just a product name.
  • Good: "Founders comparing project management tools" or "People frustrated with slow CI pipelines looking for advice"
  • Weak: "AcmeApp" or "software"

The contextual filter is what stops keyword matches from flooding you with irrelevant threads. A post that mentions "project management" in a meme or off-topic rant should score lower than a genuine recommendation request.

Filtering strictness (Loose / Medium / Tight) adjusts how picky this step is. Tight = embeddings must agree more strongly that the thread fits your focus.

2. Suggest subreddits (community matching)

When you click Suggest subreddits during Listener setup, Earshot embeds your focus text and searches a catalog of Reddit communities for semantically similar ones.

What this means for you:

  • Richer focus text → better suggestions. "B2B SaaS founders discussing fundraising and churn" will surface different subs than "startups" alone.
  • Suggestions are starting points — always review before adding. Earshot matches meaning, not your personal familiarity with a community.
  • If suggestions feel off, rewrite your focus text with more specific angles (audience, problem, intent) and click Suggest subreddits again.

3. Knowledge Bases (draft-time retrieval)

When you upload documents to a Knowledge Base, Earshot splits them into sections and embeds each chunk. At draft time — when Solo generates a reply for a matched thread — it embeds the thread context and pulls the most relevant chunks from your attached KBs.

What this means for you:

  • Accurate, specific documents beat long marketing PDFs. Embeddings find passages that mean the same thing as the question, not pages that merely contain a keyword.
  • One topic per section or file helps. A pricing FAQ and an integrations list are easier to match than one giant unstructured dump.
  • Stale content = wrong drafts. Embeddings will happily retrieve outdated pricing if that's what's in the file. Keep KBs current.

You don't need to tag chunks or add special formatting. Plain, clear writing in .md or .txt is enough.

4. Persona reply examples (voice matching)

On Solo, if your Persona includes reply examples, Earshot can use embeddings to compare draft context against those examples and prefer replies that sound like your past writing.

What this means for you:

  • Add real reply examples that show how you actually talk — not landing-page copy.
  • Examples that match the kinds of threads you target help more than one generic "thanks for sharing!" line.

This is subtler than focus text or KB retrieval, but it's the same idea: similar meaning (and tone) → better fit.

What embeddings don't do

  • They don't replace keywords. You still need primary keywords. Embeddings refine matches; they don't scan all of Reddit alone.
  • They don't read your mind. Vague focus text produces vague matches. Be specific about intent and audience.
  • They don't guarantee factual replies. KB retrieval helps, but you always review drafts before posting.
  • They aren't visible in the UI. You won't see embedding scores on draft cards today. You'll see the outcome: relevant threads, useful KB facts in drafts, and the contextual filter stage in the health funnel.

What's next