Why does searching my call transcripts for a word miss calls that obviously mention it?
Because callers rarely use your word. They describe a symptom, name a brand, or say something the recognizer transcribed slightly differently. Keyword search matches strings; it has no idea that a leaking water heater and no hot water are the same problem. Semantic search matches meaning and catches those, at the cost of returning some near misses. Serious searching uses both together.
Three separate reasons a keyword search fails
- Vocabulary mismatch. You search the technical term; the customer used the household one, or named a brand instead of a product category.
- Recognition substitution. The audio was fine but the recognizer wrote a similar-sounding word, so the string is simply not there to find.
- Fragmentation. The concept appears across three turns of conversation and never as a single phrase anywhere in the transcript.
What semantic search does differently
Semantic search converts passages into numerical representations of meaning and finds passages whose meaning is close to your query. It will find the household phrasing, survive a transcription substitution, and match a concept expressed across a couple of sentences.
The tradeoff is precision. Meaning-based matching returns things that are adjacent but not what you meant, and it cannot be relied on for exact-match tasks such as finding every call containing a specific confirmation code. Choosing the right tool per question is a normal part of connecting models to real business data.
The practical pattern: broad first, exact second
Start semantic to see the shape of the problem. Read a dozen results and note the actual language customers use. Then build a keyword or phrase filter from that observed vocabulary to get a countable, reproducible population.
This matters because reporting needs reproducibility. A number in a monthly report should come from a defined filter that returns the same set every time it runs, not from a fuzzy match whose boundaries drift. Semantic search is for discovery; explicit rules are for measurement.
Filters beat cleverness
Most searches people struggle with are not really search problems. They are missing-metadata problems. If you cannot narrow by date range, source, location, agent, call outcome and category before you search the text, you are asking the language layer to do work that structured fields should be doing.
Getting those fields onto every call, joined from the phone system, the ad platforms and the CRM, is the unglamorous half of call intelligence. Once they exist, a question like which paid-search calls in one market mentioned a competitor last quarter becomes a filter plus a short query rather than a research project. That is also why we tend to build the reporting surface and the search over the same underlying records.
Topics: transcript search · semantic search · recall · precision
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.