rule-based
1960s → today for safety-critical
IF matches "i want"
THEN "why do you want"
intent classifier
+ slot filling
+ state machine
pro: predictable,
deterministic,
no hallucination
con: narrow scope
doesn't handle novel inputs
retrieval-based
2010s. still ships for FAQs
encode utterance,
find nearest match,
return stored response
threshold for refusal
pro: handles paraphrase,
no hallucination,
scales cheaply
con: can't combine info,
can't do multi-turn
reasoning
neural (seq2seq)
2014-2019. rarely solo now
encoder-decoder on
conversation logs
generates from scratch
pro: fluent phrasing
con: generic outputs
("I don't know"),
factual drift, no grounding,
no tool use
used inside hybrids today,
not solo
LLM agent loop
2023 → 2026 production
plan → call tool
→ observe → decide
→ ... → answer
RAG grounding
tool contracts
eval + observability
pro: takes actions,
handles novelty
con: harder to reason about,
prompt injection,
infinite loops