# Exact Phrase vs. Keyword vs. Semantic Search

> Choose between phrase, keyword, and semantic search based on what you remember. Compare useful clues, limitations, and when to change approach.

- URL: https://www.hemory.com/blog/searchable-memory/phrase-keyword-semantic-search/
- Language: English (en)
- Author: Hemory Team
- Published: 2026-11-04
- Last updated: 2026-11-04
- Summary: A comparison of exact phrase, keyword, and semantic retrieval for recovering a detail from a past conversation. It explains how to choose from reliable recall clues, adapt to the actual mismatch, and interpret the resulting passage. An invoice-approval example preserves the customer's constraint. Hemory contributes captured context and supported agent access without an asserted search-mode interface.
- Tags: conversation search, exact phrase search, keyword search, semantic search

You need to check what a customer said about invoice approval. You remember the problem—payments were held up when a manager was away—but not the customer's words. Searching for a polished phrase such as “approval bottleneck” may overlook a discussion about invoices waiting in someone's inbox.

The best starting point depends on what you remember. Use an exact phrase for reliable wording, keywords for distinctive terms, and meaning-based search for an idea whose wording is uncertain. This guide explains the choice and how to change approach when the results do not answer your question.

## TL;DR

- Match the search to your strongest clue. There is no required exact-phrase → keyword → semantic sequence.
- Exact phrase search helps locate known wording; keywords preserve useful names and terms; semantic search helps bridge different descriptions of the same idea.
- A tool may combine these approaches. A natural-language search box does not reveal how it retrieves results.
- Judge the passage against your task, including the customer, scope, and statement's meaning. A literal match can still be irrelevant.
- If the conversation was captured in [Hemory](/features/), its transcript and organized Memory provide context to revisit without first reconstructing the discussion yourself.

## Compare the three approaches by the clue you have

“Semantic search” is the more usual term for search based on meaning; “topic search” can also mean browsing a topic label. The distinction matters when you decide what to try.

| Approach | Useful starting clue | What it helps you find | What still needs attention |
|---|---|---|---|
| Exact phrase | Wording you remember accurately, such as `backup approver` | Occurrences of that wording under the tool's phrase-matching rules | The phrase may occur in a suggestion, rejection, or unrelated project |
| Keyword | A distinctive term, name, or identifier | Material containing relevant terms, with matching behavior set by the tool | Common terms can match different issues; required terms can exclude useful passages |
| Semantic | A remembered idea, such as invoices waiting while a manager is away | Passages discussing a related meaning, potentially with different words | Similar meaning does not establish the same customer, condition, or conclusion |

These are retrieval approaches, not necessarily three buttons in an app. For example, [Microsoft's hybrid search documentation](https://learn.microsoft.com/en-us/azure/search/hybrid-search-overview) describes combining keyword and vector retrieval in one request. That is an example of search design, not a description of Hemory's internal implementation.

You can use this comparison without knowing an engine's architecture. Look at the available controls and the results: does the tool find the wording, a related idea, or both?

## Choose a starting point without discarding a useful clue

If you remember a distinctive phrase, use it. Replacing `backup approver` with a broad request about financial efficiency may remove the detail that distinguishes this conversation from dozens of others.

If you remember the issue but not the phrase, describe the issue immediately. There is no benefit in inventing a sentence and treating it as an exact quote. “Invoices waiting while the manager is away” is a better starting description than a guessed line in quotation marks.

If you have both kinds of clue, keep both. A query about `Northstar` and invoices waiting for an absent manager preserves the customer clue while describing the problem. Use the search behavior your tool supports; typing a name into a question is not the same as applying a verified speaker filter.

Confidence also matters. If you are sure about the customer but only guessing the month, start with the customer. An overly narrow date restriction can hide the right conversation. If the known transcript is already open, reading the relevant section may be quicker than changing search modes.

## Adjust to the mismatch you actually see

Use the result to decide which part of the query needs changing.

| What happened | Useful adjustment | Why it helps |
|---|---|---|
| A remembered phrase finds nothing | Try its distinctive word or describe the idea | The recorded wording may differ |
| A broad topic returns several customers' discussions | Add the reliable customer or project clue | The topic is right, but the scope is wrong |
| A code produces similar-looking identifiers | Look for literal occurrences and inspect the complete code | The identifier itself matters |
| A strict query requires several uncertain words | Remove the least reliable requirement | One guessed term may exclude the passage |
| The right conversation appears, but the snippet is incomplete | Open the discussion around the match | Another search may add less value than the surrounding exchange |

Do not assume quotation marks, punctuation, or multiple words behave identically everywhere. [Elastic's phrase-query documentation](https://www.elastic.co/docs/reference/query-languages/query-dsl/query-dsl-match-query-phrase), for example, describes analyzed terms and configurable spacing between them. A phrase search is not universally a character-for-character comparison.

Likewise, a keyword query may require all terms, accept some, or apply additional language processing. A failed query for two words does not tell you which rule was used. Follow the product's documented behavior when a particular operator matters.

## Example: recover the concern behind an approval request

This fictional interview illustrates alternative search routes; it is not a benchmark or actual Hemory result:

> Interviewer: What happens when your finance manager is away?
>
> Customer: The invoices sit in her inbox until she gets back. We need a backup approver, but whoever approves must still be on the finance team.

Different remembered clues suggest different starting requests:

- If you remember `backup approver`, look for that phrase using supported phrase search.
- If you remember `invoices` and the customer, use those distinctive clues.
- If you remember only the delay, describe invoices waiting for an absent manager through a meaning-based search or conversation Q&A option.

These are alternative routes to inspect, not reported search results. Once you locate the exchange, the useful answer is:

> The customer needs a backup approver so invoices do not wait for the finance manager's return. Approval must remain within the finance team.

Finding the phrase alone would not establish the condition. Finding a semantically similar discussion about speeding up payments would not establish this customer's request. The task is complete when the recovered passage supports both the problem and the relevant constraint.

## Use captured conversations without rebuilding them first

Start with the records you already have: an interview transcript, meeting notes, a customer record, or a captured conversation. A complete existing note may answer the question directly.

If you captured the exchange with Hemory, its transcript preserves the discussion and its organized Memory helps you revisit the relevant context. Its built-in Q&A can retrieve across captured conversations. You can ask:

> What delayed invoice approval, and what restriction did the customer place on a backup approver? Show the relevant passage.

Begin with the question and use the source passage to check the returned meaning. You do not need a list of extracted quotes first.

If the next task is a requirements draft across several captured interviews, a compatible AI agent connected to Hemory can retrieve from the Hemory context available to it and draft the relevant section. Give it the customer and the output you need. Supply any necessary material outside that accessible context, such as the current requirements document. The benefit is less manual selection and retelling before the AI can help.

For future interviews, capturing the exchange also lets you focus on follow-up questions. You do not have to decide in the moment which exact expression will be useful later.

## FAQ

### Is semantic search always less precise than exact phrase search?

No. A broad phrase can appear in many irrelevant passages, while a well-specified meaning-based request can locate the right issue. Phrase matching constrains wording; it does not establish relevance to the task. Compare the actual results, not a universal ranking of methods.

### Should I keep searching after I find the relevant passage?

Only when the task requires more. Recovering a customer's wording may be complete once the passage and attribution are clear. Establishing the current requirement may also require checking later accepted changes. Let the output determine the stopping point.

### What if the source is missing?

Check another available record or ask the person involved a focused question. A missing result alone does not establish that the exchange was never captured. If no usable record exists, a later recollection can help with the work, but it is a new account of the earlier discussion.

Start with the clue you trust most. Use the retrieved context to answer the work question, and change approach only when the current result leaves something material unresolved.

For the matching rules behind a phrase, see [what exact phrase search means](/blog/searchable-memory/what-is-exact-phrase-search/). For a practical identifier lookup, use [finding exact phrases, names and codes](/blog/searchable-memory/find-exact-phrases-in-conversations/).