pea-audit

Audit French PEA (Plan d'Épargne en Actions) eligibility of ETFs by reading their KID (Key Information Document) with a vision LLM. Tells you whether a fund is actually eligible for a French PEA account — with verbatim citations from the document.

What is a PEA? France's tax-sheltered stock account (€150k cap, gains tax-free after 5 years). It only accepts EU-domiciled equities, or UCITS funds that synthetically replicate non-EU indexes (S&P 500, MSCI World, Nasdaq, …) via a swap on an EU-equity basket. Physical-replication funds of non-EU indexes — most iShares Core / Vanguard ETFs — don't qualify. This library tells you which side of that line your fund is on.

What's in this repo? Two things: pea-audit — the library you pip install (lives in pea_audit/) — and ETFTracker — a reference app that consumes it (Streamlit dashboard + CLI + FastAPI at the repo root, plus etftracker/ helper code). Most of this README is about the library; see ETFTracker.md for the app side (French).

Latest: v0.2.0 — async API, typed Enums, prompt-version cache key, SSRF guard, 51 unit tests. Full history in CHANGELOG.md.

$ python audit_cli.py samples/amundi_pea_monde_kid.pdf
📄 Audit de : samples/amundi_pea_monde_kid.pdf

  ✅ ÉLIGIBLE PEA    (confiance : high)

  Émetteur     : Amundi
  ISIN         : FR001400U5Q4
  Indice       : MSCI World Index EUR
  Réplication  : synthetic_swap

  Le fonds est éligible au PEA car il utilise une réplication synthétique
  via swap (IFT) avec un panier d'actions européennes ≥75%.

  Preuves :
    p.1 — « Le Fonds est éligible au Plan d'Épargne en Actions français (PEA) ... »
    p.1 — « La performance sera échangée contre celle de l'Indice de Référence ... »

Why

PEA eligibility is opaque and changes silently — issuers re-domicile, swap counterparties, switch to ESG-screened variants, and rename funds (e.g. Amundi PEA Nasdaq-100 silently became "Amundi PEA US Tech Screened" under the same ticker). Brokers don't always flag this. pea-audit reads each fund's KID directly and tells you what the document actually says, with quotes you can verify.

Install

pip install pea-audit

Optional extras:

pip install 'pea-audit[observability]'  # adds Langfuse for LLM tracing
pip install 'pea-audit[evals]'           # adds pyyaml for the eval suite
pip install 'pea-audit[dev]'             # everything above + python-dotenv

Quickstart

Get an Ollama Cloud key at https://ollama.com/settings/keys, then:

from pathlib import Path
from pea_audit import audit_pdf, VerdictCache
from pea_audit.llm import OllamaCloudClient

# Ollama Cloud keys look like "<32-hex-char id>.<24-char secret>"
# (not "sk-..." — that's the OpenAI format)
llm = OllamaCloudClient(api_key="abcdef0123456789abcdef0123456789.EXAMPLE-KEY-DO-NOT-USE")

# Cache is opt-in. Library never writes to disk unless you supply one.
cache = VerdictCache(Path("./cache"))

verdict = audit_pdf("path/to/kid.pdf", llm=llm, cache=cache)

print(verdict.eligible)        # Eligible.YES | NO | UNCERTAIN  (also == "yes" / "no" / "uncertain")
print(verdict.replication)     # Replication.PHYSICAL | SYNTHETIC_SWAP | UNKNOWN
print(verdict.isin)            # deterministic — extracted from PDF text + Luhn-validated
for c in verdict.evidence:
    print(f"  p.{c.page}: « {c.quote} »")

Don't have a KID PDF handy? The repo ships samples/amundi_pea_monde_kid.pdf — clone or download it to try the example end-to-end on a real (PEA-eligible) Amundi fund.

More examples? See examples/ — 5 runnable demos (basic audit, audit-by-ticker, custom VisionLLM, custom KIDSource, async batch with asyncio.gather).

Audit by ticker (built-in URL registry)

from pea_audit import audit_ticker, VerdictCache
from pea_audit.llm import OllamaCloudClient

llm = OllamaCloudClient(api_key="<your-ollama-cloud-key>")
cache = VerdictCache(Path("./cache"))

result = audit_ticker("EWLD.PA", llm=llm, kid_dir=Path("./kids"), cache=cache)
print(result.verdict.eligible)  # Eligible.YES

Built-ins ship for the most common French ETFs (Amundi PEA range, BNP Paribas Easy). Add more:

from pea_audit.sources import register_source, KIDSource

register_source(KIDSource(
    ticker="LYX.PA",
    isin="FR0010411884",
    url="https://www.lyxoretf.fr/.../kid.pdf",
    issuer="Lyxor",
))

Async usage

Use aaudit_pdf + AsyncOllamaCloudClient from asyncio code — FastAPI handlers, webhook receivers, parallel batches:

import asyncio
from pea_audit import aaudit_ticker, VerdictCache
from pea_audit.llm import AsyncOllamaCloudClient

async def main():
    llm = AsyncOllamaCloudClient(api_key="...")
    cache = VerdictCache(Path("./cache"))
    # 4 audits in parallel, ~one HTTP round-trip total instead of four
    return await asyncio.gather(*[
        aaudit_ticker(t, llm=llm, kid_dir=Path("./kids"), cache=cache)
        for t in ["EWLD.PA", "PAEEM.PA", "ESE.PA", "PANX.PA"]
    ])

asyncio.run(main())

AsyncVisionLLM is the protocol sibling of VisionLLM — bring your own async provider (Claude vision via anthropic, OpenAI, …).

Architecture

flowchart LR
    A[KID PDF] --> B[pypdfium2<br/>rasterize pages]
    A --> C[pypdfium2<br/>text layer]
    C --> D[ISIN regex<br/>+ Luhn check]
    B --> E[VisionLLM<br/>analyze_images]
    D -.-> E
    E --> F[PeaVerdict<br/>eligible / replication<br/>isin / evidence]
    F --> G[VerdictCache<br/>sha256-keyed]
    G --> H[Your app:<br/>CLI / Streamlit / FastAPI / …]

    style E fill:#dbeafe,stroke:#1e40af
    style D fill:#dcfce7,stroke:#166534
    style F fill:#fef3c7,stroke:#854d0e

The LLM judges what the document says; deterministic regex + Luhn reconciles the ISIN string (vision is fuzzy on alphanumerics). The cache is opt-in — pass cache=None for a stateless library.

Two protocols make it extensible without forking:

VisionLLM — swap the model

from typing import Any, Protocol

class VisionLLM(Protocol):
    def analyze_images(
        self,
        images: list[bytes],
        prompt: str,
        schema: dict[str, Any],
        system: str | None = None,
    ) -> dict[str, Any]: ...

The default OllamaCloudClient wraps Gemma 4 via Ollama Cloud with tenacity retries on transient errors and optional Langfuse tracing. Anyone can implement this protocol to plug in Claude vision, GPT-4o, Gemini, a local Ollama instance, etc. An AsyncVisionLLM sibling exists for async backends.

KIDSource — add issuers

from pea_audit.sources import register_source, KIDSource, get_source, all_sources

A registry of ticker → KID URL mappings. Ships builtins for Amundi (URL pattern), BNP Paribas (per-fund UUIDs); URL helpers for BlackRock/iShares + Vanguard are importable but don't auto-register (most of their funds are PEA-ineligible — they're for testing the negative path).

Eval baseline

The repo ships 13 regression cases under evals/cases/*.yaml — 7 PEA-eligible synthetic-swap, 6 ineligible physical non-EEA — covering Amundi, BNP, BlackRock/iShares, Vanguard. Current baseline on Gemma 4 31b-cloud: 13/13 (100%). Run before any prompt or model change:

python evals/run.py                  # compares against evals/baseline.json
python evals/run.py --save-baseline  # snapshot the current pass-set

Exit code 2 on regression (a previously-passing case now fails) — wire into CI to gate prompt/model changes.

What does it cost?

Default backend is Gemma 4 31b-cloud via Ollama Cloud:

Bring your own LLM via VisionLLM and the cost equation becomes your provider's per-image price × ~3 pages per KID. The library doesn't add overhead beyond one model call per audit.

Production niceties

• Retries on transient errors — tenacity with exponential backoff (1s → 4s → 16s), only on network/timeout/5xx (not on 4xx or schema errors that won't self-resolve)
• SSRF guard on downloads — _download_kid rejects non-http(s) schemes, enforces a 20 MB streaming cap, verifies Content-Type looks like PDF before writing to disk (matters because KIDSource.url is user-registrable)
• Optional observability — Langfuse traces per LLM call (model, input/output, tokens, latency). Activates when LANGFUSE_PUBLIC_KEY/LANGFUSE_SECRET_KEY are set, silent no-op otherwise
• Deterministic ISINs — vision misreads of the 12-char ISIN string are corrected by regex-extracting candidates from the PDF text layer and validating with the Luhn check digit
• Versioned prompts — pea_audit/prompts/audit_v{N}.md files, selected via prompt_version= parameter; rollback is a config change, not a code edit. The version is part of the cache key, so upgrading a prompt automatically invalidates stale verdicts
• Hard vs soft fields in diffs — compare_verdicts() defaults to comparing only categorical fields (eligible, replication, isin) so monthly re-audit doesn't false-fire on LLM rephrasing of free-text issuer/index names

Known limitations

• LLM variance — across repeated runs with the same prompt + model + PDFs, the eval pass rate oscillates between 11/13 and 13/13. The two flaky cases (iShares Core S&P 500 + iShares Nasdaq 100) sometimes return replication: unknown instead of physical; the LLM's summary_fr still reasons correctly, only the structured field wavers. Eligibility verdict is stable in all observed runs. Candidate v0.3 fix: multi-sample voting (run N=3, take majority).
• Vision-only ISIN reads can drift — mitigated by the Luhn-validated text-layer extractor, but scanned PDFs without a text layer fall back to the LLM's vision read and can be wrong on alphanumeric ISINs (e.g. FR001400U5Q4 → FR00140056U4). The verdict & replication fields are reliable; the ISIN field on scanned PDFs is best-effort.
• The audit verdict is LLM-judged, not regulatory advice. The library cites the actual KID text so you can verify — always do, especially before buying.

Reference app: ETFTracker

The repo also ships a personal-tool app that consumes the library: a French ETF portfolio tracker with a Streamlit dashboard, monthly re-audit cron, FastAPI service, and Docker compose deployment. See ETFTracker.md (French) for that side.

To run it:

cp positions.csv.example positions.csv   # edit with your own holdings
cp .env.example .env                     # add your OLLAMA_API_KEY
docker compose up -d web                 # → http://localhost:8502
# or:  streamlit run dashboard.py

Streamlit "Portefeuille" tab — 4 holdings with live yfinance prices and PEA-eligibility badges (✅ from the audit cache).

Not a developer?

Three options if you don't write Python but want to check your PEA:

1. Run the dashboard locally with the 3 commands above. No code edits required after positions.csv is filled.
2. Upload via the HTTP API — docker compose up -d api then POST /audit/upload with a PDF (Swagger docs at http://localhost:8080/docs).
3. Hire a developer — realistically the best option for non-technical PEA holders. The library exists so a hosted version of this is buildable in a weekend.

Contributing

See CONTRIBUTING.md. Maintainer? See PUBLISHING.md. Release history: CHANGELOG.md.

License

MIT.