Should I embed fund factsheets in a vector store or call an API at query time?

Why embeddings are the wrong tool for factsheet figures

A factsheet is mostly tables: holdings, sector weights, calendar returns, cost lines. Chunking it for a vector store breaks rows from headers, so "5.48" ends up near "NVIDIA" by luck. Fund houses publish a new version every month, so the store fills with near-duplicate chunks from different dates and the retriever cannot tell which is current. And the same fund has several share classes with separate documents whose text differs by one letter in the name; similarity search returns the wrong one routinely.

The result is confident answers with stale or mismatched numbers, which in a financial product is worse than no answer.

What works: structured data as a tool

Give the model a function that takes an ISIN and returns the current JSON. The payload is already structured (headlineMetrics.ter, topHoldings[], sector[], calendarReturns, dataAsOf), so the model reads exact values with a date and never has to parse a PDF.

python
# pip install fundfacts
from fundfacts import FundFacts
ff = FundFacts()  # reads FUNDFACTS_API_KEY
​
def fund_context(isin: str) -> str:
    """Compact, dated context block for the prompt."""
    f = ff.get_fund(isin)
    d = f["data"]
    lines = [f"{f['name']} ({isin}), data as of {d['dataAsOf']}",
             f"Category {d['profile']['category']}; risk {d['riskRating']}/7; TER {d['headlineMetrics']['ter']}; size {d['keyFacts']['aum']}",
             "Top holdings: " + ", ".join(f"{h['name']} {h['weight']}%" for h in d["topHoldings"][:10]),
             "Sectors: " + ", ".join(f"{s['label']} {s['weight']}%" for s in d["sector"][:6]),
             "Calendar returns: " + ", ".join(f"{y} {r}%" for y, r in zip(d["calendarReturns"]["years"], d["calendarReturns"]["fund"]))]
    return "\n".join(lines)

Inject that block into the prompt (or expose it as a tool and let the model call it), and keep the vector store for the material that genuinely is unstructured: your research, meeting notes, client correspondence.

Where documents still matter

For a fund the API does not cover, POST /extract (Pro and above) turns any KID or factsheet PDF into the same data shape, so your application handles both cases with one code path instead of falling back to raw text.

Guardrails worth adding

• Instruct the model to quote dataAsOf with every figure and to say "not disclosed" for empty fields.
• Validate ISINs before calling (12 characters, check digit) so the model cannot invent one that happens to parse.
• Cache by ISIN until expiresAt; the figures do not change more than once a day.

The fund data for AI agents and LLMs guide covers evaluation and prompt patterns in more depth.