valiantlynxz/tripletex-tool-embeddings

Tripletex API Tool Embeddings

Pre-computed embeddings for 800 Tripletex accounting API tools, extracted from the OpenAPI 3.0.1 spec and embedded with Google gemini-embedding-001 (3072 dimensions).

Built for RAG-based tool filtering in the AI Accounting Agent competition project.

Quick Start

from datasets import load_dataset

# Full embeddings (800 tools, 3072-dim vectors) — ready for RAG
ds = load_dataset("valiantlynxz/tripletex-tool-embeddings")

# Lightweight: tool metadata only, no embedding vectors
ds = load_dataset("valiantlynxz/tripletex-tool-embeddings", name="tools")

Configurations

Config	Default	Columns	Size	Use case
embeddings	Yes	name, description, parameters, embedding	~9 MB	RAG search, vector index
tools		name, description, parameters	~50 KB	Browsing, filtering, analysis

Schema

embeddings config

ds = load_dataset("valiantlynxz/tripletex-tool-embeddings")
example = ds["train"][0]

example["name"]         # "AccountantDashboardNews_get"
example["description"]  # "Get public news articles"
example["parameters"]   # '{"from": "integer", "count": "integer", ...}'  (JSON string)
example["embedding"]    # [3072 floats] — gemini-embedding-001 vector

tools config

ds = load_dataset("valiantlynxz/tripletex-tool-embeddings", name="tools")
example = ds["train"][0]

example["name"]         # "AccountantDashboardNews_get"
example["description"]  # "Get public news articles"
example["parameters"]   # '{"from": "integer", "count": "integer", ...}'  (JSON string)

Data Summary

• 800 API tools from the Tripletex accounting API (OpenAPI 3.0.1)
• 3072-dimensional embeddings via Google gemini-embedding-001
• Parameters stored as JSON strings mapping param names to types
• Source: openapi.json included in this repo (3.5 MB, 546 paths, 2167 schemas)

Using with LanceDB

from datasets import load_dataset
import lancedb

ds = load_dataset("valiantlynxz/tripletex-tool-embeddings")

# Convert to LanceDB
db = lancedb.connect(".tool_embeddings")
records = [
    {
        "name": row["name"],
        "description": row["description"],
        "parameters": row["parameters"],
        "embedding": row["embedding"],
    }
    for row in ds["train"]
]
table = db.create_table("tools", data=records, mode="overwrite")

# Search
results = table.search(query_embedding).limit(100).to_list()

Using with FAISS

from datasets import load_dataset
import numpy as np
import faiss

ds = load_dataset("valiantlynxz/tripletex-tool-embeddings")

embeddings = np.array(ds["train"]["embedding"], dtype=np.float32)
index = faiss.IndexFlatIP(3072)
faiss.normalize_L2(embeddings)
index.add(embeddings)

# Search
query = np.array([query_embedding], dtype=np.float32)
faiss.normalize_L2(query)
distances, indices = index.search(query, k=100)
tool_names = [ds["train"][int(i)]["name"] for i in indices[0]]

Regenerating Embeddings

The scripts/ directory contains the original embedding pipeline:

• scripts/embeddings.py — Google Gemini embedding provider
• scripts/rag_tool_filter.py — OpenAPI-to-embedding pipeline + LanceDB vector store

# Requires: google-genai, lancedb
# Requires: GCP_API_KEY environment variable

from scripts.embeddings import get_embedding_provider
from scripts.rag_tool_filter import ToolEmbedder, ToolVectorStore, index_openapi_tools
import json, asyncio

with open("openapi.json") as f:
    spec = json.load(f)

provider = get_embedding_provider()
embedder = ToolEmbedder(provider)
store = ToolVectorStore(".tool_embeddings")
asyncio.run(index_openapi_tools(spec, store, embedder))

Repo Structure

tripletex-tool-embeddings/
├── README.md
├── embeddings.parquet      # 800 tools with 3072-dim embeddings
├── tools.parquet           # 800 tools metadata only (lightweight)
├── openapi.json            # Source Tripletex OpenAPI 3.0.1 spec (3.5 MB)
└── scripts/
    ├── embeddings.py       # Google Gemini embedding provider
    └── rag_tool_filter.py  # OpenAPI extraction + LanceDB indexing

Source Project

Part of the nmai project — ai-accounting-agent/.