Documentation

VectorFin is designed as a drop-in retrieval layer for Retrieval-Augmented Generation (RAG). Embed the user's question with the same model we use (gemini-embedding-2-preview), call POST /v1/embeddings/search, then pass the top chunks to your LLM with citations.

import os, requests, google.generativeai as genai

genai.configure(api_key=os.environ["GEMINI_API_KEY"])
VF_KEY = os.environ["VECTORFIN_API_KEY"]

question = "What did Apple management say about services margins in 2024?"

# 1. Embed the query with the same model VectorFin uses.
q_vec = genai.embed_content(
    model="gemini-embedding-2-preview",
    content=question,
    task_type="retrieval_query",
)["embedding"]

# 2. Retrieve the top-k chunks (the "R" in RAG).
hits = requests.post(
    "https://api.vectorfinancials.com/v1/embeddings/search",
    headers={"X-API-Key": VF_KEY},
    json={"query_vector": q_vec, "ticker": "AAPL", "top_k": 8},
).json()["data"]

# 3. Ground the LLM with citations (the "AG" in RAG).
context = "\n\n".join(
    f"[{h['ticker']} {h['fiscal_period']} chunk {h['chunk_idx']}] {h['text']}"
    for h in hits
)
answer = genai.GenerativeModel("gemini-2.0-flash").generate_content(
    f"Answer using ONLY the context below, citing [ticker fiscal_period chunk_idx].\n\n"
    f"Context:\n{context}\n\nQuestion: {question}"
)
print(answer.text)

Want point-in-time RAG (e.g. answer Q2 2022 questions without future information)? Pass knowledge_ts_max in the search body — the bitemporal index filters out anything we learned after that date.