Deep Search

Search audio by meaning. Find what was said even when you don’t know the exact words.

Imagine you’re scanning 50 episodes of a health podcast looking for the best supplements. You don’t know the product names yet. That’s the whole point. You can’t search for “Shilajit” or “Tongkat Ali” because you’ve never heard of them. What you can describe is what you’re looking for: “supplement that actually improved energy,” “what I take every morning,” “the one thing I’d recommend for men over 30.”

Keyword search is useless here. You’d have to already know the answers to search for them. deep_search flips this: you describe the idea, and Augent finds the exact moments where products, recommendations, and protocols are discussed, even when you have no idea what they’re called. The names, the dosages, the timestamps, all surfaced from a natural language prompt.

---

Searching by meaning

deep_search doesn’t match words. It matches meaning.

Under the hood, it converts every segment of the transcript into a numerical fingerprint that captures its meaning. Your query gets the same treatment. Augent finds the segments whose meaning is closest to yours and ranks them by relevance.

Describe what you’re looking for in plain language and Augent finds the moments where that idea was discussed, regardless of the exact words.

Query: “supplements that improve energy and testosterone”

Finds:

• “The data on Shilajit is really compelling for men over 30” — ep-33, 34:22
• “Most people would benefit from Tongkat Ali” — ep-11, 1:12:05
• “If you’re feeling burned out, this is the first thing I’d add to your stack” — ep-07, 48:15

None of these results necessarily contain the words “energy” or “testosterone.” All of them are exactly what you were looking for.

---

Real example: supplement research across 50 episodes

Take 50 episodes of a health podcast and run deep_search with queries like “supplement that actually works,” “what I take every morning,” and “the one thing I’d recommend.”

What you get back would take weeks to compile by hand:

Supplement	Benefit	Who It’s For	Daily	Episode	Timestamp
Shilajit	Testosterone, energy	Men 30+	Yes	ep-33	34:22
Tongkat Ali	Hormone optimization	Men, athletes	Yes	ep-11	1:12:05
Ashwagandha	Stress, cortisol regulation	High-stress individuals	Cycling	ep-29	28:41
Apigenin	Sleep onset	Anyone with sleep issues	Yes	ep-42	52:07
Oregano Oil	Gut health, antimicrobial	Immune support	Yes	ep-15	18:33
Turmeric	Anti-inflammatory, joints	General health, athletes	Yes	ep-22	41:09

With matched quotes and timestamps:

“The data on Shilajit is really compelling for men over 30” — ep-33, 34:22

“Most people would benefit from Tongkat Ali” — ep-11, 1:12:05

“The literature strongly supports Ashwagandha for stress” — ep-29, 28:41

“If you have trouble falling asleep, Apigenin is the one to look at” — ep-42, 52:07

Fifty hours of audio. Every recommendation pulled out, sourced, and timestamped. Minutes, not weeks.

---

Same pattern, any domain

The supplement example is one use case. Same approach works anywhere you have hours of content and don’t know exactly what to search for:

Domain	What you’d query	What gets surfaced
Product research	”tool they swear by,” “product that saved them time”	Specific software, hardware, and services you’ve never heard of, with firsthand endorsements and context
Competitive intel	”feature their customers love,” “what they do differently”	Positioning, strategy, and advantages mentioned in interviews, demos, and earnings calls
Legal & compliance	”admitted fault,” “agreed to terms,” “changed their story”	Critical moments in depositions, hearings, and recorded calls
Hiring & culture	”why they left,” “what made them stay,” “red flag in the interview”	Patterns across exit interviews, all-hands recordings, and candidate screens
Education	”the key insight,” “most common mistake,” “what I wish I knew”	Actionable takeaways from lectures, courses, and tutorials without watching the full thing
Content repurposing	”best quote,” “most controversial take,” “funniest moment”	Clip-worthy moments across an entire content library

The workflow is always the same: describe what you’re looking for, get back the exact moments with names, details, and timestamps that you couldn’t have found with keywords alone.

---

Finding what you didn’t know to search for

This is what deep_search is built for.

Keyword search only works when you already know what you’re looking for. But the most valuable content in any recording is often something you didn’t expect: a passing recommendation, an offhand insight, a product mention you’d never heard of. You can’t write a keyword query for something you don’t know exists.

With semantic search, broad queries surface specific moments:

Query	What it finds
”product that changed their routine”	A specific brand or supplement the host credits with a measurable result
”tool they use every day”	Software, hardware, or workflow the speaker can’t live without
”something most people get wrong”	Contrarian insights and corrections buried in long-form conversation
”underrated strategy that actually works”	Tactics and approaches that aren’t mainstream but have proven results

These aren’t keyword matches. They’re meaning matches. The speaker never said “underrated strategy.” They told a story about what worked. But the meaning aligned, and deep_search found it.

Instead of listening through hours of audio, you mine across it. You’re not searching for words. You’re searching for ideas — and finding ones you didn’t know were there.

---

Embeddings are shared and stored

The first time deep_search runs on a file, it computes embeddings for every segment and stores them in memory. Every subsequent search on that file is instant, no matter how different the query. The expensive computation happens once.

These embeddings are also shared with chapters and search_memory (semantic mode), which use the same vectors. Run deep_search on a file, and chapters or search_memory on that file is free. The embeddings are already computed.

Need to search across all stored transcriptions at once? Use search_memory — it queries every file in memory with a single call, no audio_path needed.

---

When to use deep search vs. keyword search

Both tools exist because they solve different problems:

	Keyword search	Deep search
Best for	Product names, brand names, proper nouns, specific phrases	Concepts, themes, ideas, exploratory queries
Query	"Ashwagandha"	"supplement that helps with stress"
Matches	Exact word occurrences	Semantically similar segments
Speed	Instant (string matching)	Instant after first run (embedding lookup)
Misses	Anything said differently	Nothing meaningful. If the idea is there, it finds it

Use search_audio when you know the exact term: a product name, a brand, a person. Use deep_search when you know the idea but not the words, or when you want to discover what’s in the content without knowing in advance.

Tool Reference

Parameters, response format, and technical details