Semantic search or keyword matching: which tool fits?
Word overlap and embedding similarity answer different questions. Neither is a fact checker.
When exact words matter
Use Keyword in context to inspect literal occurrences, or Keyword coverage to check supplied phrases. These tools can show where a phrase appears without assuming what the sentence means. A negative statement can still contain your target words.
When related wording matters
The semantic tools encode English sentences with MiniLM. Search ranks original sentences against a question; comparison returns a cosine score. Similar wording may be related even when it does not answer the question. Scores are relative similarities, not probabilities.
Know the model limits
The current embedding model accepts English only. Source-based tasks support at most 40 sentences, each at most 800 characters. Every model input is also checked against 256 tokenizer tokens; overlong text is rejected rather than silently truncated.
Keep a reference to the source
Model summaries extract original sentences. Grouping places source sentences in clusters, and topic matching ranks the labels you provide. Review the cited text, try a more precise query and avoid treating a cluster or top label as a verified fact.