knowlytix-benchmark

Benchmark for Structured Retrieval from Financial Documents. Auto-generates questions from document-graph topology and scores LLM predictions against a provably correct graph-traversal baseline. (Internally referred to as the FinStructBench module — knowlytix.benchmark.*.)

knowlytix-benchmark is one of four packages in the Geometric Memory Systems family. Use it to answer questions like "how close does my RAG pipeline get to the graph-verified ground truth on this financial report?" — with metrics that resist gaming because the ground truth comes from graph operations, not a held-out human-labeled test set.

• Package: knowlytix-benchmark
• License: Apache-2.0
• Python: 3.12+
• Status: alpha (v0.x)

Install

pip install knowlytix-benchmark

Depends on knowlytix-core (pinned ~=0.1.0). LLM-mode scoring routes through LiteLLM — set GMS_LLM_MODEL plus your provider's API key and any model works.

Quickstart — score a prediction set

import json
from importlib.resources import files

from knowlytix.benchmark import score_answer

# Smoke fixtures shipped with the wheel:
questions = json.loads((files("knowlytix.benchmark.fixtures.smoke") / "questions.json").read_text())
predictions = json.loads((files("knowlytix.benchmark.fixtures.smoke") / "predictions.json").read_text())

by_id = {p["id"]: p["answer"] for p in predictions["predictions"]}
for q in questions["questions"]:
    result = score_answer(by_id[q["id"]], q["ground_truth"])
    mark = "correct" if result.correct else "wrong"
    print(f"{q['id']}: {mark}  partial={result.partial_score:.2f}  ({result.detail})")

Quickstart — run the full benchmark

from knowlytix.benchmark import Benchmark, get_instance_path

bench = Benchmark(get_instance_path("model_validation"))
result = bench.run()          # graph-only mode (no LLM, no API key needed)
bench.print_results(result)

To evaluate an LLM against the same ground-truth graph:

from knowlytix.benchmark.llm_caller import create_client

client = create_client()       # reads GMS_LLM_MODEL_SCORER → GMS_LLM_MODEL
result = bench.run(llm_client=client)
bench.print_results(result)

CLI

benchmark --instance model_validation
benchmark --instance credit_portfolio --llm-model anthropic/claude-opus-4-6

Configuration

FINSTRUCTBENCH_* — scoring tolerances

GMS_LLM_* — LLM routing (only needed for LLM-mode scoring)

See .env.example in the source repo for the full provider key reference.

Included benchmark instances

Five synthetic financial-domain instances ship with the wheel:

All synthetic — no real institution, person, or market event is depicted.

Public API

from knowlytix.benchmark import (
    Benchmark, BenchmarkResult,
    DocumentGraph, ENMEntry, ENMKey, PhaseEncoder,
    FinStructBenchSettings,
    GeneratedQuestion, ScoreResult, score_answer,
    default_generators, get_instance_path, ingest_markdown, list_instances,
)

GeneratedQuestion is a stable contract consumed by knowlytix.harness.testing.bridge — don't rename without coordinating (see CLAUDE.md §Coding standards).

Related packages

Links

• Source: knowlytix/gmsh
• Paper: FinStructBench: Benchmarking Structured Retrieval from Financial Documents