Battery-cell specifications as structured, semantic, linked data, and a GraphRAG workflow that answers questions about the cells from the knowledge graph instead of from an LLM's memory.
- 382 cell profiles (JSON-LD) in
BatteryTypeJson/, typed with the EMMO battery domain ontology: rated capacity, voltages, currents, cycle life, mass, dimensions, case and cathode chemistry, plus DOI links to publications that use or reference each cell. batterygraph, a small Python package that loads the profiles into an RDF graph (RDFLib), provides query helpers, and runs the GraphRAG workflow with OpenAI, Anthropic, Google Gemini or local Ollama models.- Evaluation: a 300-question benchmark with recorded results for three models with GraphRAG and one LLM-only baseline.
BatteryTypeJson/ 382 cell profiles (JSON-LD)
batterygraph/
graph.py load the graph, EMMO term <-> IRI mapping, query helpers
llm.py one interface for OpenAI, Anthropic, Gemini and Ollama
prompts.py query-generation, answer and grading prompts
rag.py GraphRAG: question -> helper calls / SPARQL -> answer
evaluate.py grading, "identifiably wrong" check, summary, figures
notebooks/
KnowledgeGraphDemo.ipynb SPARQL and helper-function examples
GraphRAG.ipynb ask questions, small evaluation, recorded figures
scripts/
run_eval.py run the benchmark for a model (with or without RAG)
make_figures.py recreate figures and summary from recorded results
generate_qa_dataset.py generate a Q&A benchmark from the profiles
evaluation/
qa_dataset.csv the 300 questions and reference answers
results/ recorded answers and grades, summary.csv
figures/ RAG_Accuracy.svg, RAG_HallucinationRobustness.svg
pip install -r requirements.txt # or: pip install -e ".[all]"Only the packages for the LLM providers you use are needed. The knowledge graph itself needs rdflib, requests and pandas.
API keys. Copy .env.example to .env and fill in the keys you need (OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY). .env is git-ignored. Local models via Ollama need no key.
Offline use. The profiles reference the context https://w3id.org/emmo/domain/battery/context, which is downloaded once when the graph loads. To work offline, save context.json and set BATTERYGRAPH_CONTEXT=/path/to/context.json.
from batterygraph import BatteryGraph, GraphRAG
bg = BatteryGraph("BatteryTypeJson")
bg.query_cell_property("INR21700-50E", "RatedCapacity", manufacturer="Samsung SDI")
# [{'cell': 'Samsung SDI INR21700-50E', 'property': 'RatedCapacity', 'value': 4.9, 'unit': 'Ah'}]
bg.get_cell_info("LF280K") # manufacturer, case, cathode material, linked publications
# SPARQL with readable EMMO terms in {braces}
bg.sparql("""
SELECT (COUNT(DISTINCT ?cell) AS ?n) WHERE {
?cell {hasPositiveElectrode}/{hasActiveMaterial} ?m . ?m a {LFP} .
}""")
rag = GraphRAG(bg, model="claude-sonnet-5")
rag.answer("Which battery has a higher rated capacity: the Saft VL 5U or the Lishen LR2170SA?")- The LLM receives the question, the graph structure, the property names and the manufacturer list, and writes calls to three helper functions (
query_cell_property,get_all_cell_properties,get_cell_info) or, for questions across many cells, a SPARQL query with{Term}placeholders. - The output is run against the graph. Helper calls are parsed, not evaluated: only these three functions with literal arguments are executed, and SPARQL must be read-only (
SELECT/ASK). - The LLM answers from the query result only. If the graph has no data, the answer should say so rather than guess.
The benchmark (evaluation/qa_dataset.csv) covers single-property lookups, multi-property summaries, comparisons, case and chemistry questions, derived values (Wh, mAh) and negative controls (made-up cells, missing specifications). Each answer was graded against the reference answer by the answering model itself; whether an incorrect answer admits missing data was judged by gemini-3.5-flash for all runs (a few unparseable grades were also re-graded by gemini-3.5-flash).
| Model | Mode | Accuracy | Hallucination robustness |
|---|---|---|---|
| gemini-3.5-flash | GraphRAG | 0.81 | 0.66 |
| claude-sonnet-5 | GraphRAG | 0.79 | 0.24 |
| gpt-5.4 | GraphRAG | 0.76 | 0.54 |
| gemini-3.5-flash | LLM only | 0.42 | 0.01 |
Accuracy = correct answers / all questions. Hallucination robustness = answers that admit missing data ("I don't know") / incorrect answers.
Recreate the figures and summary.csv from the recorded results (no model calls):
python scripts/make_figures.pyRun the benchmark yourself:
python scripts/run_eval.py --check gemini-3.5-flash # test API access
python scripts/run_eval.py --model claude-sonnet-5 # GraphRAG
python scripts/run_eval.py --model gemini-3.5-flash --no-rag # LLM-only baselineThe recorded results were produced with an earlier version of the profiles and of the query prompt. A new run on the current data and code will give somewhat different numbers.
Each file in BatteryTypeJson/ describes one cell as JSON-LD with the EMMO battery domain context:
@typeBatteryCell,schema:name(model) andschema:manufacturer(with a Wikidata link where known)hasProperty: numerical properties such asRatedCapacity,NominalVoltage,CycleLifeorMass, each with its value, unit (where applicable) and aschema:citationof its sourcehasCase: case / form factor (e.g.PrismaticCase,R21700);hasPositiveElectrode/hasActiveMaterial: cathode chemistryschema:subjectOf: publications that mention the cell, identified by DOI only (3,930 links to 3,406 papers for 176 cells).schema:additionalTypeholds this project's classification of each link:USED,REFERENCEDorUNCLEAR(154 links are unlabelled).
Identifiers. 358 cells have a persistent battinfo IRI (https://w3id.org/battinfo/spec/…) as @id. The other 24 are not yet registered there and carry a local UID, which JSON-LD resolves against the document URL.
BSD 3-Clause License, see LICENSE.