MetacouplingLLM: An LLM-Powered Python Package for Applying the Metacoupling Framework

1. Overview

MetacouplingLLM (import name metacouplingllm) is a Python package (v0.2.1) that helps researchers apply the telecoupling and metacoupling frameworks (Liu et al., 2013; Liu, 2017) to their own research topics using Large Language Models (LLMs). Given a natural-language research description, the package produces a structured, framework-compliant analysis that identifies coupled human and natural systems, classifies coupling types, maps flows, agents, causes, and effects, and grounds the output in both a curated literature database and real-time web search results.

Key capabilities:

• Structured metacoupling analysis from free-text research descriptions
• Multi-turn refinement via conversational LLM interaction
• Retrieval-Augmented Generation (RAG) over 421 telecoupling/metacoupling papers — 219 indexed as full text, the other 202 as structured paraphrased summaries (used wherever redistributing the full text would not be licence-safe)
• RAG-only literature Q&A mode for users already familiar with the framework (coupling_analysis=False)
• Literature recommendation from a curated BibTeX database
• Real-time web search grounding with native backends for OpenAI, Anthropic, Gemini, and Grok (DuckDuckGo fallback for custom clients and whenever a native backend fails or returns nothing)
• evidence_coverage_note: the model's closing self-assessment (§7 Evidence Coverage, requested in every framework-mode analysis) of how well the retrieved literature passages and, when web search is on, the web results support each group of claims
• Pericoupling validation against two geographic databases (country-level and subnational ADM1) with mode-aware presentation: region-scale adjacency for subnational queries, country-scale for national queries; spillover pairs filtered (left to the framework systems analysis)
• Automated map generation at both country and subnational levels, including supranational-union dissolution (EU / ASEAN / USMCA)
• Support for OpenAI, Anthropic, Google Gemini, xAI Grok (each adapter auto-wires its native web search), or any custom LLM backend (web search falls back to DuckDuckGo)
• Scholar-ready export: result.abstract, result.to_markdown(), result.to_docx() for manuscript-ready output
• Quantitative metacoupling indicators (metacouplingllm.indicators) — IFS / PFS / TFS, MFE, MFCI computed on user-supplied flow data
• Optional LLM-assisted indicator helpers — define_study, check_inputs, classify_ambiguous_edges, interpret_results, write_methods, each returning (result, LLMTrace) for reproducibility
• Built-in run tracing (on by default) — every analyze()/refine() records a RunTrace (the prompts, responses and token usage of every model call made through the assistant's client, plus the git SHA) and writes an inspectable artifact folder under runs/ (the main-analysis, map-extraction and abstract calls, raw web results, retrieved chunks, parsed analysis, map data and the formatted report); the web-stage model calls (native web search, when used, and structured web extraction) bypass the traced client, so only their results are kept; disable with trace=False

Requirements: Python 3.10+, two hard runtime dependencies: numpy>=1.21 and fastembed>=0.8 (local ONNX embeddings: the corpus vectors ship with the package, but fastembed downloads the ~0.2 GB BAAI/bge-base-en-v1.5 model on first use, so the first retrieval needs network access or a populated model cache). Optional extras: openai, anthropic, gemini, grok, viz (geopandas + matplotlib), search (ddgs), indicators (pandas), export (python-docx), all.

Two complementary tracks

The package exposes two complementary analysis tracks that can be used independently or combined:

1. Deterministic core — the national (ADM0) and subnational (ADM1) adjacency databases and the metacouplingllm.indicators submodule, which computes flow-share / evenness / concentration indicators on user-supplied flow data. Five optional LLM-assisted helpers (define_study, check_inputs, classify_ambiguous_edges, interpret_results, write_methods) wrap natural-language judgment tasks around the deterministic core.
2. Qualitative LLM-driven case-study analysis via MetacouplingAssistant — produces a structured, framework-compliant report with citations, optional maps, and scholar-ready Markdown / Word export.

Quick Install

pip install metacouplingllm
pip install "metacouplingllm[openai]"
pip install "metacouplingllm[anthropic]"
pip install "metacouplingllm[gemini]"
pip install "metacouplingllm[grok]"
pip install "metacouplingllm[search]"
pip install "metacouplingllm[viz]"
pip install "metacouplingllm[indicators]"   # quantitative indicators (pandas)
pip install "metacouplingllm[export]"       # scholar export (python-docx)
pip install "metacouplingllm[all]"

If you want to use the built-in OpenAI example below, install metacouplingllm[openai].

What Users Get

• A structured metacoupling analysis from a plain-language study description
• Optional literature recommendations and supporting evidence passages
• Optional web-grounded context for current trade, policy, or event information
• A closing evidence_coverage_note in every framework-mode analysis: the model's self-assessment of which retrieved literature passages and web sources back which claims
• Optional map generation for country-level and ADM1-level analyses
• Scholar-ready abstract + Markdown + Word export (result.abstract, result.to_markdown(), result.to_docx())
• Quantitative metacoupling indicators (IFS / PFS / TFS, MFE, MFCI) computed on user-supplied flow data via metacouplingllm.indicators
• Optional LLM-assisted helpers for study setup, input validation, ambiguous-edge classification, results interpretation, and methods-text drafting — each producing reproducible LLMTrace records

---

2. Theoretical Foundations

The package operationalizes the metacoupling framework (Liu, 2017), which integrates three types of human-nature interactions across geographic scales:

Coupling Type	Definition	Example
Intracoupling	Interactions within a single coupled human and natural system (CHANS)	Manure management impacts on local water quality in Michigan
Pericoupling	Interactions between adjacent coupled systems	Feed grain trade between Michigan and neighboring Indiana
Telecoupling	Interactions between distant coupled systems	Pork exports from Michigan to Japan

The framework identifies five core components in each coupling:

1. Systems -- Sending, receiving, and spillover systems, each with human and natural subsystems
2. Flows -- Movements of matter, capital, energy, information, people, and organisms
3. Agents -- Fixed categories: individuals / households; firms / traders / corporations; governments / policymakers; organizations / NGOs; non-human agents
4. Causes -- Fixed categories: economic; political / institutional; ecological / biological; technological / infrastructural; cultural / social / demographic; hydrological; climatic / atmospheric; geological / geomorphological
5. Effects -- The same fixed categories used for causes, applied to outcomes across coupled systems

The package encodes 14 telecoupling categories from the literature (trade, migration, tourism, species invasion, water transfer, etc.) and the six-phase operationalization procedure from Liu (2017).

---

3. Architecture

                    Research Description (free text)
                              |
                    +---------v----------+
                    |  MetacouplingAssistant   |
                    +---------+----------+
                              |
          +-------------------+-------------------+
          |                   |                   |
  +-------v-------+  +-------v--------+  +-------v-------+
  | PromptBuilder  |  |  Web Search    |  | RAG Engine    |
  | (6-layer       |  |  (native +     |  | (421 papers,  |
  |  system prompt)|  |   DuckDuckGo)  |  |  embeddings   |
  |                |  |                |  |  or TF-IDF    |
  |                |  |                |  |  fallback)    |
  +-------+-------+  +-------+--------+  +-------+-------+
          |                   |                   |
          +-------------------+-------------------+
                              |
                    +---------v----------+
                    |     LLM Client     |
                    |  (OpenAI / Anthropic /|
                    |   Gemini / Grok /   |
                    |   custom backend)   |
                    +---------+----------+
                              |
                    +---------v----------+
                    |    Parser          |
                    |  (structured       |
                    |   ParsedAnalysis)  |
                    +---------+----------+
                              |
          +-------------------+-------------------+
          |                   |                   |
  +-------v-------+  +-------v--------+  +-------v-------+
  | Coupling       |  | Literature     |  | Map Generator |
  | Validation     |  | Recommendations|  | (ADM0/ADM1)   |
  +---------------+  +----------------+  +---------------+
                              |
                    +---------v----------+
                    |   AnalysisResult   |
                    |  .parsed           |
                    |  .formatted        |
                    |  .abstract         |
                    |  .to_markdown()    |
                    |  .to_docx()        |
                    |  .map              |
                    +--------------------+

Quantitative indicators side-track:

       User-supplied flow DataFrame
                       |
            +----------v-----------+
            |  build_adjacency     |  ← fills the adjacent flag
            |  (optional)          |     from the bundled ADM0 /
            |                      |     ADM1 databases (or supply
            |                      |     your own adjacency table)
            +----------+-----------+
                       |
            +----------v-----------+
            |  classify_coupling   |  ← assigns I / P / T
            |  (optional LLM       |     to each edge
            |   fallback)          |
            +----------+-----------+
                       |
            +----------v-----------+
            |  compute_flow_shares |  ← IFS / PFS / TFS
            |  compute_mfe         |  ← Shannon evenness
            |  compute_mfci        |  ← normalised HHI
            |  summarize_metacoupling
            +----------+-----------+
                       |
            +----------v-----------+
            |  LLM helpers         |  ← define_study,
            |  optional; each      |     check_inputs,
            |  returns (result,    |     classify_ambiguous_edges,
            |   LLMTrace)          |     interpret_results,
            |                      |     write_methods
            +----------------------+

Indicator math never calls an LLM (deterministic-first by design). The LLM helpers are scoped to natural-language and judgment tasks (study setup, input validation, ambiguous-edge classification, interpretation, methods drafting) and always pair their output with an LLMTrace for reproducibility. classify_ambiguous_edges, and classify_coupling(llm_client=...) which calls it, can assign the I / P / T labels the indicators are computed from, so those labels may vary between runs, while the indicator math itself stays deterministic.

---

4. Core Functions

4.1 MetacouplingAssistant

The central class that orchestrates the entire analysis pipeline.

from metacouplingllm import (
    JOURNAL_ARTICLES_2025,
    MetacouplingAssistant,
    OpenAIAdapter,
)
from openai import OpenAI

advisor = MetacouplingAssistant(
    llm_client=OpenAIAdapter(OpenAI(api_key="..."), model="gpt-4o"),
    auto_map=True,              # Generate map automatically
    rag_corpus=JOURNAL_ARTICLES_2025,  # Use bundled 2025 journal corpus
    web_search=True,            # Ground analysis in web search results
    web_search_max_results=10,  # Number of web results (default)
    web_structured_extraction=True,  # the default; validated countries + flows
    rag_top_k=8,                # Number of RAG evidence passages (default)
    rag_min_score=0.60,         # embeddings backend only; omit for the backend-aware default
    max_examples=2,             # Framework examples in prompt
    temperature=0.7,            # LLM temperature
    verbose=True,               # Print progress messages
)

Key methods:

Method	Purpose
analyze(research_description)	First-turn structured analysis
refine(additional_info, focus_component=None)	Multi-turn follow-up refinement
reset()	Clear conversation for a new topic

4.2 Prompt Engineering (6-Layer System)

The system prompt is constructed in six layers:

1. Role -- Expert persona in metacoupling and sustainability science
2. Knowledge -- Full framework definitions, 14 categories, coupling transformations
3. Methodology -- Six-phase operationalization procedure (Liu, 2017)
4. Examples -- The best keyword-matching published case studies (up to max_examples) from five bundled examples: Brazil-China soybean trade, red imported fire ant invasion, Beijing water supply, panda conservation / SDGs, Wolong Nature Reserve
5. Output Format -- Structured template for seven analysis sections (with a Citation Rules sub-layer, 5b: turn-scoped [Tk:N]/[Tk:Wn] citation mechanics and hallucination guardrails)
6. Interaction -- Multi-turn refinement guidelines

Before calling the LLM, the system conditionally injects (each only fires if its precondition is met):

• Pericoupling-database hints — at most one of two flavors fires per query:
  • Country-level hint: if ≥1 country name is detected and no ADM1 region applies. If ≥2 countries are named, each focal-vs-other pair is reported with its database classification (e.g., "Mexico (MEX) and United States (USA) are pericoupled" for adjacent pairs, "Brazil (BRA) and China (CHN) are telecoupled" for distant pairs). If exactly 1 country is named, the focal country's full pericoupled-neighbor list is injected as candidate context (e.g., a Mexico-only query gets Belize, Guatemala, USA presented as REFERENCE — the LLM is encouraged to consider these as potential pericoupled flows but must confirm with independent evidence). Framing is REFERENCE-only in both cases.
  • ADM1 (subnational) hint: if any ADM1 region (e.g., Michigan, Mato Grosso) is detected. The focal region's adjacent neighbors are listed as reference-only adjacency rather than asserted pericoupling — the LLM is told that adjacency alone is NOT evidence of pericoupling and must find independent flow evidence.
  • Queries with no detectable country (and no ADM1 region) get no hint; queries with one country named but no pericoupled neighbors in the database (island nations like Australia, Japan, Cuba) also get no hint.
• Web search context with [Tk:W1], [Tk:W2] labels for inline citation (turn-scoped — k is the conversation turn). Only when web_search=True and at least one hit returned.
• Structured web map hints — receiving/spillover countries and flows extracted by a second LLM pass over the web results, each kept only if it resolves to a real country/union and cites a real retrieved web result (confidence ≥ 0.7). This grounds the evidence, not the spillover role (which classification the LLM still decides). Only when web_structured_extraction=True (the default).

4.3 Retrieval-Augmented Generation (RAG)

The RAG engine provides evidence grounding from a corpus of 421 telecoupling and metacoupling papers. The split is by what can be redistributed, not by topic: 219 papers whose licences allow it are indexed as full text, and the remaining 202 are indexed as structured, paraphrased summaries (metadata plus key findings, no source text) so they stay searchable without reproducing copyrighted material. Those 202 are papers the corpus cannot redistribute in full — 125 paywalled articles and 77 articles not cleared for redistribution (free to read without a redistribution licence, restrictively licensed, available only as a preprint, or with open-access status not established):

• Indexing: Papers are chunked by section and indexed by one of two backends:
  • Embeddings (default) -- semantic retrieval via fastembed + the BAAI/bge-base-en-v1.5 ONNX model. Captures synonyms, paraphrases, and related concepts (e.g., a query about "soybean trade" also matches chunks about "soya bean exports" and "Glycine max shipments"). Pre-computed corpus vectors are shipped with the package as chunk_embeddings.npy (~15 MB) so users never have to re-encode.
  • TF-IDF (fallback) -- lexical retrieval using TF-IDF + cosine similarity. Activated when fastembed is unavailable or the pre-computed file is missing.
• Retrieval: Cosine similarity with a per-paper cap (default 3 chunks/paper, configurable via rag_max_chunks_per_paper; within-paper section diversity preferred). The relevance floor is backend-aware — rag_min_score defaults to 0.60 for embeddings (BGE-base cosine) and 0.01 for TF-IDF (the two score scales are not comparable).
• Citation: Evidence passages are appended as [Tk:1], [Tk:2], ... with inline annotation (turn-scoped — k marks the turn so prior-turn references remain unambiguous across multi-turn conversations)
• Lightweight: fastembed + onnxruntime add ~20 MB to the install; no torch/GPU dependencies
• Backend selection: MetacouplingAssistant(..., rag_backend="auto") (default) picks embeddings if available and transparently falls back to TF-IDF. Explicit options: "embeddings", "tfidf".

4.4 Web Search Grounding

Web search injects current, real-world context (trade data, policies, recent events) that may not be in the LLM's training data:

• Native auto-wiring when web_search=True: each adapter routes to its own backend — OpenAIWebSearchBackend, AnthropicWebSearchBackend, GeminiWebSearchBackend (Google Search grounding), GrokWebSearchBackend (xAI's server-side web_search tool on the /responses endpoint). Custom clients, and any native backend that raises or returns no results, fall back to the built-in DuckDuckGo search cascade.
• DuckDuckGo backend has three fallback layers: ddgs -> duckduckgo_search -> stdlib (urllib + html.parser); works on Google Colab without any extra packages.
• Results cited as [Tk:W1], [Tk:W2], ... -- the W prefix distinguishes web sources from literature [Tk:1], [Tk:2]. k is the turn index, so prior-turn web references stay stable across refine() calls.
• On by default (web_structured_extraction=True): a second LLM pass (strict JSON / tool-output across all four providers) over the web results and validates map-ready countries and flows before using them in auto-maps.
• evidence_coverage_note is the model's 2-5 paragraph closing self-assessment (§7), requested in every framework-mode analysis, of how well the retrieved literature passages and web sources support each group of claims and which claims rest on inference — surfaced on result.parsed.evidence_coverage_note and result.evidence_coverage_note.
• Supranational unions (EU / ASEAN / USMCA / NAFTA) are handled end-to-end: the map-data extraction prompt lets the model name a union as a flow target and the package expands it to its member states, web-result summaries are passed to the map-data extraction call up to 2,500 characters each, and the map renderer dissolves union borders and draws the bloc as one region, filled as pericoupling if any member borders the focal country and as telecoupling otherwise.

4.5 Pericoupling Databases

Two geographic adjacency databases validate LLM coupling classifications:

Database	Scope	Coverage
Country-level	Countries and territories	Full global (ISO alpha-3)
ADM1 (subnational)	First-level administrative regions	3,374 regions, 8,462 shared-border pairs (8,026 under the default moderate coupling standard), 196 countries

Functions: is_pericoupled(), get_pericoupled_neighbors(), lookup_adm1_pericoupling(), etc.

Both levels accept two orthogonal toggles: de_facto_borders (default True; fold disputed land into its de-facto administrator, vs the strict WB standard layer) and coupling_standard (stringent / moderate / lenient, default moderate) — for pairs sharing only a river/lake border, moderate requires a completed fixed crossing (road or railway bridge, causeway, dam-top road or tunnel; ferries and crossings under construction do not count, while a finished bridge on a politically closed border or a disused railway bridge with its track in place does), stringent drops all water-only pairs, and lenient keeps them. Each fixed-crossing flag was adjudicated by the same two-model design as the water-only status (an OpenStreetMap screen, GPT-5.6 Sol research, Sonnet-5 adversarial judgment, a location test and maintainer map rulings); see docs/METHODS_adjacency.md §8.

Region names are resolved to ADM1 codes by resolve_adm1_code, which—beyond unaccented forms and native-script letters that NFKD cannot decompose—consults a bundled, deterministically-validated English-exonym alias table (1,145 aliases; e.g. Bavaria→Bayern, Tuscany→Toscana) and returns None rather than guessing on an ambiguous or padded name. (The prompt-hint scanner and the validator's text cleaner additionally strip possessive suffixes, e.g. "Michigan's"→"Michigan", and split hyphenated compounds, e.g. "Michoacán-Jalisco", in free text before calling it; resolve_adm1_code itself does neither — resolve_adm1_code("Michigan's") returns None even though resolve_adm1_code("Michigan") returns "USA023".)

4.6 Literature Recommendations

From a curated BibTeX database of 265 empirical telecoupling/metacoupling journal articles (2013–2026), the system ranks papers by author keywords (3.0 points per match), title words (2.0 per match), full-text relevance from the RAG index (up to 3.0) and a log-scaled citation bonus (up to 2.0), with ties broken by citation count and then year. Call get_database_info() for live counts if the corpus drifts.

4.7 Map Visualization

Three map functions generate matplotlib figures:

Function	Level	Data Source
plot_focal_country_map(focal_country)	Country	Local World Bank Admin 0 all_layers if available; otherwise hosted Admin 0_all_layers mirror; fallback official World Bank Admin 0 10m
plot_analysis_map(parsed_analysis)	Country	Local World Bank Admin 0 all_layers if available; otherwise hosted Admin 0_all_layers mirror; fallback official World Bank Admin 0 10m
plot_focal_adm1_map(focal_adm1)	Subnational (ADM1)	Cached World Bank Admin 1 if available; otherwise hosted Admin 1.gpkg mirror, with World Bank NDLSA.gpkg for disputed-area overlay

Map features: coupling-colored regions, flow arrows, disputed territory hatching, customizable color palettes.

4.8 LLM Client Abstraction

The package uses a protocol-based design that supports any LLM backend:

from metacouplingllm import (
    AnthropicAdapter,
    GeminiAdapter,
    GrokAdapter,
    LLMResponse,
    OpenAIAdapter,
)

# Built-in adapters, each wrapping its provider's own SDK client. openai_client,
# anthropic_client, genai_client and xai_client stand for clients you create yourself:
# openai.OpenAI(...), anthropic.Anthropic(...), google.genai.Client(...), and for Grok
# an openai.OpenAI(...) created with base_url="https://api.x.ai/v1"
OpenAIAdapter(openai_client, model="gpt-4o")
AnthropicAdapter(anthropic_client, model="claude-sonnet-4-20250514")
GeminiAdapter(genai_client, model="gemini-2.5-flash")
GrokAdapter(xai_client, model="grok-3")     # OpenAI-protocol-compatible

# Any custom client with a chat() method also works
class MyClient:
    def chat(self, messages, temperature=0.7, max_tokens=None):
        return LLMResponse(content="...")

Each built-in adapter auto-wires its native web-search backend when web_search=True; custom clients, and any native backend that raises or returns no results, fall back to DuckDuckGo.

4.9 Scholar Export

AnalysisResult exposes three exporter surfaces so users can turn an LLM analysis into something a co-author or reviewer can read:

Surface	Description	Install needed
result.abstract	One-paragraph scholar abstract from a separate, final LLM call (prompted not to add details absent from the analysis; skip it with generate_abstract=False)	none
result.to_markdown(path=None)	Manuscript-ready Markdown with auto-derived title, bold sub-field labels, per-category Flows / Causes / Effects subsections, RAG + web evidence blocks	none
result.to_docx(path)	Word document with the same structure as the Markdown export; a multi-bullet §1 classification is rendered as bullet paragraphs	metacouplingllm[export] (python-docx)

4.10 Quantitative Indicators

The metacouplingllm.indicators submodule computes deterministic metacoupling indicators on user-supplied flow data:

Function	What it does
classify_coupling(edges, focal_id, adjacency, ...)	Add coupling_type (I / P / T) to an edge table using a user-supplied adjacency table (or one generated by build_adjacency).
build_adjacency(pairs, level="adm0", ...)	Fill the adjacent flag (1 / 0 / <NA> when a name cannot be resolved) from the bundled country ("adm0") or ADM1 ("adm1") pericoupling database, with the same de_facto_borders / coupling_standard options; pass the result to classify_coupling.
compute_flow_shares(data, ...)	Intracoupled / Pericoupled / Telecoupled Flow Shares (IFS / PFS / TFS) per focal system.
compute_mfe(data, ...)	Metacoupled Flow Evenness — normalised Shannon entropy across coupling types; data must be the shares table returned by compute_flow_shares (columns IFS / PFS / TFS, set by share_cols), not the raw edge table.
compute_mfci(data, ...)	Metacoupled Flow Concentration Index — normalised HHI (IFCI / PFCI / TFCI) within each coupling type.
summarize_metacoupling(data, ...)	One-shot combined indicator table.

Built on established statistics (Shannon 1948 entropy; Hirschman 1945 HHI, normalised per Hannah & Kay 1977; Equivalent Number of Partners per Laakso & Taagepera 1979) and adapted to the metacoupling framework as Metacoupled Flow Shares (MFS), Evenness (MFE) and Concentration Index (MFCI); the package introduces no new statistic. Brazil-soybean worked example: IFS = 0.10, PFS = 0.20, TFS = 0.70, MFE ≈ 0.73, TFCI ≈ 0.33. See MANUAL §16 for the full math + code.

Pandas is the only added dependency; install via pip install "metacouplingllm[indicators]".

4.11 LLM-Assisted Indicator Helpers

Five optional helpers in metacouplingllm.indicators wrap natural-language judgment tasks around the deterministic indicator core. Each returns (result, LLMTrace) for reproducibility (see MANUAL §17, LLM-Assisted Indicator Helpers):

Helper	What it does
define_study(description, *, llm_client)	Natural-language description → structured study config dict.
check_inputs(data_summary, sample_rows, *, llm_client)	Validate user data; flag missing inputs, unit issues, self-loop intracoupling.
classify_ambiguous_edges(edges, study_config, *, llm_client)	Classify edges the deterministic pass couldn't resolve; returns "I" / "P" / "T" / "unknown" per edge with confidence and reason.
interpret_results(results, *, llm_client, audience)	Plain-language interpretation of an indicator table; audience presets "academic" / "general" / "policy".
write_methods(indicator_spec, *, llm_client)	Manuscript-ready Methods text with formulas + standard citations (Shannon 1948, Hirschman 1945, Hannah & Kay 1977, Laakso & Taagepera 1979, Liu 2017).

The LLMTrace dataclass carries timestamp_utc, model, prompt_version, system_prompt, user_prompt, raw_response, usage so old traces stay attributable when prompt wording evolves.

Automatic LLM resolution of ambiguous edges in classify_coupling(): pass llm_client= (plus optional study_config=) and the function automatically calls classify_ambiguous_edges() on any rows the deterministic pass left as NaN, merging suggestions back. A model= keyword runs the call on another model: a built-in adapter is copied with that model (its other settings kept), and a custom client receives model= on chat(). When the LLM returns "unknown", the row stays NaN; any "I" / "P" / "T" label it does give is written into coupling_type whatever its confidence, with the call's LLMTrace kept in the returned table's .attrs["llm_classify_trace"], so review the labels (classify_ambiguous_edges() called directly also returns each suggestion's confidence and needs_user_confirmation).

All five helpers work across OpenAI / Anthropic / Gemini / Grok adapters and use each provider's native strict-JSON / tool-output mode for the structured-output helpers.

---

5. Operation Procedure

Step 1: Install

pip install "metacouplingllm[all]"          # full installation
# or selectively:
pip install "metacouplingllm[openai]"       # OpenAI support
pip install "metacouplingllm[anthropic]"    # Anthropic support
pip install "metacouplingllm[gemini]"       # Google Gemini support
pip install "metacouplingllm[grok]"         # xAI Grok support
pip install "metacouplingllm[viz]"          # maps (geopandas + matplotlib)
pip install "metacouplingllm[search]"       # web search (ddgs)
pip install "metacouplingllm[indicators]"   # pandas-based quantitative indicators
pip install "metacouplingllm[export]"       # python-docx scholar export

Step 2: Initialize

from openai import OpenAI
from metacouplingllm import (
    JOURNAL_ARTICLES_2025,
    MetacouplingAssistant,
    OpenAIAdapter,
)

client = OpenAI(api_key="your-key")
advisor = MetacouplingAssistant(
    llm_client=OpenAIAdapter(client, model="gpt-4o"),
    auto_map=True,
    rag_corpus=JOURNAL_ARTICLES_2025,
    web_search=True,
    web_search_max_results=10,
    web_structured_extraction=True,
    rag_top_k=8,
)

Step 3: Analyze

result = advisor.analyze("""
    My research examines what the impact of Brazil's soybeans exports is.
""")

print(result.formatted)   # Full formatted report

if result.map:
    result.map.savefig("metacoupling_map.png", dpi=150, bbox_inches="tight")

The output includes:

• Coupling classification (intracoupling / pericoupling / telecoupling)
• Systems identification (focal, sending, receiving and spillover systems, each with human and natural subsystems)
• Flows analysis (matter, capital, information, energy, people, organisms)
• Agents grouped by the fixed five-category agent vocabulary
• Causes grouped by the fixed cause/effect category vocabulary
• Effects grouped by the fixed cause/effect category vocabulary
• Research gaps and suggestions
• Literature evidence with [Tk:1]-[Tk:N] citations (turn-scoped)
• Web sources with [Tk:W1]-[Tk:WN] citations (turn-scoped)
• Pericoupling database validation
• An evidence_coverage_note self-assessment from the LLM summarising what evidence backs which claims
• Scholar-ready outputs accessible via result.abstract, result.to_markdown(path), and result.to_docx(path) (the latter requires metacouplingllm[export])

Step 4: Refine (optional)

result2 = advisor.refine(
    "The main export destinations are Japan and Mexico.",
    focus_component="systems",
)
print(result2.formatted)

Step 5: Visualize (optional)

# Or generate maps independently
from metacouplingllm import plot_focal_country_map, plot_focal_adm1_map

fig = plot_focal_country_map("USA")
fig.savefig("usa_coupling_map.png", dpi=150)

fig_wb = plot_focal_country_map(
    "USA",
    adm0_shapefile=r"C:\path\to\World Bank Official Boundaries - Admin 0_all_layers.gpkg",
)
fig_wb.savefig("usa_coupling_map_world_bank.png", dpi=150)

fig2 = plot_focal_adm1_map("USA023")  # Michigan
fig2.savefig("michigan_adm1_map.png", dpi=150)

If you do not pass adm0_shapefile, the package first looks for a local World Bank Official Boundaries - Admin 0_all_layers.gpkg. If it cannot find one, it auto-downloads the hosted Admin 0_all_layers mirror into the cache on first use, then falls back to the official World Bank Admin 0 10m basemap if needed.

Similarly, if you do not pass shapefile= to plot_focal_adm1_map (adm1_shapefile= on MetacouplingAssistant), the ADM1 map looks in the cache folder (~/.cache/metacoupling/wb_adm1/, or $METACOUPLING_CACHE_DIR/wb_adm1/) and otherwise auto-downloads the hosted Admin 1.gpkg mirror into it.

Step 6: Access Structured Data

# Structured access to parsed analysis
p = result.parsed
print(p.coupling_classification)
print(list(p.iter_system_entries()))        # focal / sending / receiving / spillover system entries
#   each a dict with role, name, human_subsystem, natural_subsystem, geographic_scope;
#   sending / receiving entries also carry system_scope ("adjacent" or "distant")
print(list(p.iter_flow_entries()))          # flow dicts
print(list(p.iter_agent_entries()))         # agent dicts
print(list(p.iter_category_items("causes")))   # (coupling_type, category, item) triples
print(list(p.iter_category_items("effects")))  # (coupling_type, category, item) triples
print(p.research_gaps)        # list of research gap strings
print(p.evidence_coverage_note)  # reviewer-facing evidence self-assessment

Or hand the result to the scholar exporters:

print(result.abstract)                # one-paragraph scholar abstract
result.to_markdown("brazil_soybean.md")
result.to_docx("brazil_soybean.docx") # requires metacouplingllm[export]

Step 7 (optional): Compute Quantitative Indicators

When you have your own flow data, plug it into metacouplingllm.indicators for deterministic indicator math:

import pandas as pd
from metacouplingllm.indicators import classify_coupling, summarize_metacoupling

edges = pd.DataFrame({
    "focal_system_id":   ["Brazil"] * 6,
    "origin_id":         ["Brazil"] * 6,
    "destination_id":    ["Brazil", "Argentina", "Paraguay", "China", "EU", "USA"],
    "flow_value":        [10.0, 5.0, 15.0, 50.0, 12.0, 8.0],   # million tonnes
})
adjacency = pd.DataFrame({
    "origin_id":      ["Brazil", "Brazil"],
    "destination_id": ["Argentina", "Paraguay"],
    "adjacent":       [1, 1],
})

classified = classify_coupling(edges, focal_id="Brazil", adjacency=adjacency)
summary    = summarize_metacoupling(classified)
cols = ["focal_system_id", "IFS", "PFS", "TFS", "MFE", "IFCI", "PFCI", "TFCI"]
print(summary[cols].round(2))
#   focal_system_id  IFS  PFS  TFS   MFE  IFCI  PFCI  TFCI
# 0          Brazil  0.1  0.2  0.7  0.73   1.0  0.25  0.33

The full summary has 18 columns (focal_system_id, then F_I ... ENP_T: raw flow magnitudes, shares, MFE, the concentration indices, partner counts and equivalent numbers of partners); the snippet prints the headline indicators only.

The same tables feed the optional LLM-assisted helpers (natural-language study setup, validation, ambiguous-edge classification, interpretation, and manuscript prose around the deterministic numbers): pass the edge table (with a study_config dict) to classify_ambiguous_edges and the summary table to interpret_results; define_study takes a text description, check_inputs a summary dict (columns, row_count) plus a few sample rows, and write_methods an indicator-spec dict. Each needs llm_client= and returns (result, LLMTrace). Install via pip install "metacouplingllm[indicators]".

---

6. Bundled Data

Resource	Description
421 papers (Papers.zip)	Markdown for RAG — 219 indexed as full text, 202 as structured paraphrased summaries (papers that cannot be redistributed in full: 125 paywalled, and 77 not cleared for redistribution — free to read without a redistribution licence, restrictively licensed, preprint-only, or with open-access status not established)
Precomputed embeddings (chunk_embeddings.npy + chunk_embeddings.manifest.json)	10,970 × 768 float16 vectors (BAAI/bge-base-en-v1.5), one per chunk of the 421 papers, so the corpus never has to be re-encoded; the manifest records the chunk count and chunk-order fingerprint they were built from
BibTeX database (telecoupling_literature.bib)	265 empirical journal articles (2013–2026) with metadata for literature recommendation
Country pericoupling database (CSV)	Global country-pair adjacency classification
ADM1 edge list (CSV)	8,462 subnational shared-border pairs across 3,374 regions in 196 countries (World Bank Official Boundaries, 2026-05-14; see data/PROVENANCE.md)
ADM1 alias table (CSV)	1,145 English exonyms / alternative spellings for 863 ADM1 regions in 136 countries (e.g. Bavaria→Bayern, Tuscany→Toscana), so resolve_adm1_code matches common English names; deterministically validated (see data/PROVENANCE.md)
Water-only border files (CSV)	water_separated_pairs.csv: the 843 water-only pairs (824 ADM1 + 19 ADM0 roll-ups), each with a has_bridge fixed-crossing flag (ADM1: 388 with a crossing, 436 without), read by coupling_standard; water_classification_pairs.csv: the 824 ADM1 rows with their adjudication evidence
Disputed-territory overlay and edge-correction files (CSV)	disputed_overlay_pairs.csv: 19 pairs (16 ADM1 + 3 ADM0), read by de_facto_borders; sliver_corridor_relabel.csv (11 rows), land_gap_overlay_pairs.csv (5 pairs) and denylist_pairs.csv (7 pairs): the reviewed corrections behind the ADM1 edge list (see data/PROVENANCE.md)
Framework examples	Five curated case studies (soybean trade, red imported fire ant invasion, Beijing urban water, panda conservation, Wolong Nature Reserve); the best matches (max_examples, default 2) are injected into the prompt

Counts above are as of v0.2.1. Call get_database_info() for the live BibTeX count and len(zipfile.ZipFile(pathlib.Path(metacouplingllm.__file__).parent / "data" / "Papers.zip").namelist()) (after import zipfile, pathlib, metacouplingllm) for the live paper count if the corpus drifts.

---

7. Design Principles

• Lean dependencies -- Core analysis works with only numpy, fastembed, and an LLM client; visualization, web search, quantitative indicators, and scholar export are optional extras
• Graceful degradation -- Each optional feature (RAG, web search, maps, literature, indicators, export) can be independently enabled or disabled; RAG transparently falls back from embeddings to TF-IDF when fastembed is unavailable
• Protocol-based extensibility -- Any object with a chat() method works as an LLM client
• Pre-LLM knowledge injection, post-LLM validation -- Web search, example selection, and the pericoupling-database hint happen before the LLM call; pericoupling validation of the LLM's stated classification runs post-LLM on the parsed output and flags, in the COUPLING DATABASE VALIDATION block of result.formatted, a coupling-type mismatch between the LLM's stated classification and the database adjacency for analyst review ("Consider revising."); it never edits the LLM's text
• Semantic RAG -- Pre-computed BGE-base embeddings shipped with the package; semantic matching catches synonyms and paraphrases that TF-IDF misses. TF-IDF remains available as a fallback.
• Colab-compatible -- Web search includes a zero-dependency stdlib fallback for restricted environments
• Deterministic-first for quantitative analysis -- The metacouplingllm.indicators math never calls an LLM; numbers are reproducible from the input DataFrame alone. The five LLM-assisted helpers are scoped to natural-language and judgment tasks (study setup, input validation, ambiguous-edge classification, interpretation, methods drafting) and always pair their output with an LLMTrace record. The classification helper is instructed to return "unknown" when it cannot decide, and an "unknown" leaves the row NaN; when classify_coupling(llm_client=...) merges its suggestions, every other label (including low-confidence ones) is written into coupling_type, so review them using the confidence and needs_user_confirmation columns of classify_ambiguous_edges(); no helper changes flow values.

---

8. Testing

The package includes a comprehensive test suite. The tests are in the source repository, not in the installed wheel, so run them from a clone:

pip install -e ".[all,dev]"
pytest tests/        # CI runs: pytest tests/ -q

Tests that need pandas, python-docx or geopandas are skipped when that extra is missing ([all] installs them), and the live OpenAI tests (tests/test_openai_integration.py) are skipped unless OPENAI_API_KEY is set in the environment or a .env file, in which case they make real API calls.

1441 tests covering all modules: core advisor logic, framework enums, prompt construction, LLM parsing, RAG retrieval, literature matching, web search (including stdlib fallback), pericoupling databases, country resolution, visualization colors, map generation, scholar export, quantitative indicators, and a CI-enforced doc-capability drift guard that fails the build when shipped features aren't advertised in INTRODUCTION/README/MANUAL.