PCA tags
Tags are how Post Call Analyzer labels a call. Some tags are simple flags (“greeting said”), some extract a value (“product mentioned”), and some combine other tags into a business rule (“compliant opening”).
You configure tags on the Call Analyzer project Tags page. They run in the analysis pipeline’s apply-tags step, after transcription.
Why tags matter
Section titled “Why tags matter”QA scorecards and analytics need consistent inputs. If every reviewer invents their own labels, trends are meaningless. Tags give you a shared vocabulary:
- Supervisors filter and coach from the same flags
- QA questions can require specific tags
- Analytics can show which tags fire most often
Configure tags before you rely on them in QA conditions.
src/assets/screenshots/46-pca-tags.pngTag types
Section titled “Tag types”| Type | Idea | Best for |
|---|---|---|
| Metadata | Rules on call metadata / payload fields | Campaign, team, duration bands, CRM flags already on the row |
| BoW (bag of words) | Keyword or phrase presence in the transcript | Script phrases, product names, prohibited words |
| GenAI | Model judges or extracts from the transcript | Tone, nuance, open-ended extraction |
| Derived | Combines other tags with include / exclude logic | Composite compliance rules |
| None | Grouping only (no evaluation) | Organize a hierarchy of parent labels |
Metadata tags
Section titled “Metadata tags”Metadata tags do not read the transcript. They look at fields that arrived with the call row (payload) or pipeline output.
Example: tag vip_customer when a payload field segment equals VIP.
You choose:
- Source — payload, output, or either
- Field and operator (equals, contains, …)
- Logic — all conditions (AND) or any (OR)
Use metadata when the fact is already known before analysis. It is fast and cheap compared with GenAI.
BoW (bag of words) tags
Section titled “BoW (bag of words) tags”BoW tags scan the transcript for keywords. Matching is whole-word oriented and supports Unicode (including languages such as Hindi).
Example: keywords pricing, EMI, ब्याज for a “rate discussion” flag.
BoW is predictable and easy to audit: you can see which keyword hit. It is weaker on paraphrase (“monthly installment” might miss EMI if you only listed that token).
src/assets/screenshots/47-pca-tag-bow.pngGenAI tags
Section titled “GenAI tags”GenAI tags send the transcript (and your instructions) to the project’s LLM provider.
Two modes:
| Mode | Result | Example prompt idea |
|---|---|---|
| Flag | Yes / no style judgment | “Was the agent rude or dismissive?” |
| Extract | A value pulled from the call | “What product did the customer ask about?” |
GenAI is flexible but depends on provider quality, cost, and clear prompts. Prefer BoW or metadata when a simple rule is enough; use GenAI when judgment or free-form extraction is required.
src/assets/screenshots/48-pca-tag-genai.pngDerived tags
Section titled “Derived tags”Derived tags do not read the transcript directly. They fire based on other tags:
- Include when — require certain tag keys (all or any)
- Exclude when — suppress the derived tag if other tags matched
Example: compliant_opening includes greeting_provided and disclosure_read, and excludes prohibited_phrase.
Build base tags first, then layer derived rules. Derived tags are ideal for scorecard-friendly composites.
How to add a tag
Section titled “How to add a tag”- Open the Call Analyzer project → Tags.
- Enter a key (stable id, snake_case, for example
greeting_provided) and a human label. - Choose the type.
- Fill type-specific fields (keywords, prompt, conditions, or include/exclude rules).
- Save.
Keys should stay stable: QA and analytics refer to them. Renaming a key later is more disruptive than changing the label.
What success looks like
Section titled “What success looks like”- Each important behavior has a clear tag with the simplest type that works.
- QA questions reference tags that actually fire on real calls.
- Analyzed records show tag results that supervisors understand without reading the full transcript.
- Analytics Top tags reflects the labels you care about coaching on.
Common problems
Section titled “Common problems”| Symptom | Likely cause | What to try |
|---|---|---|
| Tag never matches | Wrong type or empty config | Confirm keywords / conditions / prompt |
| BoW misses paraphrases | Keywords too narrow | Add variants or use GenAI flag |
| GenAI inconsistent | Vague prompt | Add examples and a clear yes/no definition |
| Derived never true | Base tags inactive or keys mistyped | Match include keys exactly to base tag keys |
| QA cannot select a tag | Tag is extract-only or wrong type for flags | Use flag / BoW / derived for yes-no QA conditions |
| No tags on records | Intake not run or apply-tags failed | Check Jobs / journey; confirm runtime is up |
Tags vs QA
Section titled “Tags vs QA”Tags classify. QA scores.
A tag might say “disclosure read = yes.” A QA question might award 5 points when that tag is present, inside a bundle with other opening questions. Configure tags first, then build the QA scorecard.