For LLMs

Patterns and best practices for AI agents using HTS MCP through the Model Context Protocol.

Available Tools

HTS MCP exposes 14 MCP tools organized into 5 categories:

Category	Tools	Purpose
Tariff queries	`get_tariff`, `search_tariffs`	Look up and search tariff codes
Enrichment	`enrich_hts6_codes`, `generate_multilevel_embeddings`, `get_enrichment_status`	AI enrichment pipeline
Graph	`generate_graph_edges`, `classify_graph_edges`, `get_graph_status`, `get_code_relationships`	Knowledge graph
Resources	`search_tariff_resources`, `add_tariff_resource`, `get_resource_statistics`	Policy documents
References	`get_tariff_references`, `get_tariffs_referencing_code`, `get_database_status`	Cross-references and status

Recommended Tool Sequences

Tariff Lookup

When a user asks about a specific product:


1. search_tariffs(query="user's product description", strategy="hybrid", limit=5)
2. get_tariff(hts_code=top_result.code)           # full detail
3. get_tariff_references(hts_code=code)            # cross-references
4. get_code_relationships(hts_code=code)           # related codes

Quick Code Lookup

When the user provides an 8-digit code directly:


1. get_tariff(hts_code="0201.30.80")
2. get_tariff_references(hts_code="0201.30.80")    # optional

Exploring Relationships

When the user wants to understand connections between codes:


1. get_code_relationships(hts_code=code, min_confidence=0.7)
2. For each interesting related code: get_tariff(hts_code=related_code)

Policy Research

When the user asks about trade policy:


1. search_tariff_resources(query="topic", resource_type="guidance")
2. Present results with titles, URLs, and descriptions

Search Strategy Selection

User Intent	Strategy	Why
Natural language description	`hybrid`	Best overall accuracy, combines semantic + lexical
Specific product terms	`lexical`	Faster, exact term matching
Vague or conceptual query	`semantic`	Embedding similarity handles conceptual matches

Multi-Level Search

The level parameter controls search granularity:

Level	When to Use
`hts8` (default)	Specific tariff line lookup
`hts6`	Broader category search (uses AI-enriched descriptions)
`hts4`	Heading-level exploration
`chapter`	Broad commodity group discovery

The variant parameter controls embedding text:

Variant	Content
`full` (default)	Includes section, chapter, heading context, better for disambiguation
`short`	Code and description only, better for direct similarity

Response Interpretation

Search Results

Results include a score field (0-1). Interpret as:

0.8+: Strong match, high confidence
0.6-0.8: Good match, verify with user
0.3-0.6: Partial match, present as alternative
Below 0.3: Weak match, likely not relevant

Relationship Types

The knowledge graph classifies edges into 10 types:

Type	Meaning	Action
`material_affinity`	Same material, different forms	Note material connection
`functional_similarity`	Same purpose, different materials	Present as alternative
`substitution`	Borderline products	Warn about classification boundary
`component_assembly`	Part-to-whole	Note manufacturing relationship
`end_use`	Used together	Contextual information
`manufacturing_process`	Similar production	Background knowledge
`cross_category_bridge`	Unexpected cross-chapter link	Highlight for user
`abstraction`	Hierarchical parent-child	Navigation aid
`weak_association`	Tangential	Low priority
`no_meaningful_relationship`	False positive	Ignore

System Health

Call get_database_status at the start of a session to verify:

Tariff data is loaded
Embeddings are available (required for semantic/hybrid search)
Enrichments exist (required for HTS6-level search)
Graph data is populated (required for relationship queries)

If embeddings are unavailable, search will automatically fall back to lexical-only mode.