Plumfind’s Tariff Bot
Your smart companion for quick, accurate tariff lookups.
Important Disclaimer: This demo illustrates how retrieval and temperature settings affect AI responses. RAG-enabled answers are grounded in uploaded tariff documents, while RAG-disabled answers rely on the model’s general knowledge. Outputs are for demonstration purposes only.
Lightning Fast
Better Accuracy
Live Chat
The 4 conditions
Four configurations. One question. Four very different answers. Each cell represents a unique combination of retrieval and temperature settings.
RAG: OFF
TEMP: 0.1
Cautious General Knowledge
Model relies on training data with low creativity. Responses are hedged and conservative.
RAG: ON
TEMP: 0
Document-Grounded Precision
Retrieved documents provide exact HS codes and dates. Highly deterministic output
RAG: OFF
TEMP: 1
Broad & Variable Responses
Maximum creativity with no grounding. Risk of hallucination and imprecise figures.
RAG: ON
TEMP: 0.95
Grounded Expressiveness
Retrieval keeps facts accurate while high temperature adds richness and explanation.
Low Temperature Comparison
At low temperature, the model is deterministic and conservative. The only variable is whether documents are retrieved.
RAG: OFF
TEMP: 1
Key Observation
The model relies purely on training knowledge, producing hedged, general responses without document-backed precision.
Covers similar regional breakdowns (US, EU, China, Canada)
Ends with a disclaimer about potentially outdated information
More conservative and slightly less verbose than high-temp RAG OFF
Lacks specific HS codes or exact effective dates
Figures are approximate and drawn from general knowledge
RAG: ON
TEMP: 0
Key Observation
Document-grounded with exact HS codes and precise effective dates. The output is concise, structured, and fully deterministic.
Cites specific HS codes (e.g., 7207.11.00, 7225.99.00)
Provides exact effective date: 13-03-2025
Highly deterministic and conservative output
Tight, list-like presentation prioritizing accuracy
Sticks closely to retrieved documents without embellishment
High Temperature Comparison
At high temperature, the model becomes more expressive. The critical question: does retrieval still keep the facts accurate?
RAG: OFF
TEMP: 1
Key Observation
The model relies purely on training knowledge, producing hedged, general responses without document-backed precision.
Broad coverage of multiple regions (US, EU, China, India)
Approximate percentage ranges rather than exact rates
High verbosity and variability in output
Figures are rounded and sourced from general training data
No ability to verify claims against specific documents
RAG: ON
TEMP: 0.95
Key Observation
Retrieval keeps facts accurate while high temperature adds richness. Exact HS codes and dates remain intact with more expressive language.
Retrieves exact HS codes and precise effective dates
High temperature adds explanatory context and richer language
Response is detailed with additional breakdowns of HS codes
Grounding from RAG prevents factual drift
Best of both worlds: accuracy + expressiveness