A retrieval-augmented (RAG) research assistant over DAX 40 annual reports, with a citation guardrail and a measured evaluation harness.
Ask questions across 15 DAX company reports (FY2024 & FY2025) and get a research memo with source citations, an audit trail linking straight back to the exact page of the original filing, and year-over-year framing when both fiscal years are in scope.
The engineering claim is narrower than "it answers questions": hybrid retrieval (BM25 and dense search unioned, then cross-encoder re-ranked), a guardrail that re-parses every citation out of the generated prose and checks it against the excerpts the model was actually given before the answer is shown, retrieval and generation evals with precision@k, recall@k and a claim-level faithfulness judge, and confidence bands that stop a query with no support in the corpus from reaching the model at all. Generation sits behind a provider interface (app/provider.py), so the model vendor is configuration rather than a rewrite.
- Query 15 DAX annual reports across two fiscal years simultaneously
- Get answers with company + page citations, checked against the retrieved excerpts before they're shown — not just asserted
- Compare how companies frame the same topic differently, or how one company's own framing shifted year over year
- Filter by company, fiscal year, or query across everything at once
- Click a citation's page number to open the source PDF at that exact page
- Export any answer as a print-ready PDF memo
Example queries:
- "How do Siemens and SAP frame their AI investment strategy?"
- "Which companies disclosed structural cost reduction programs?"
- "What do CFOs say about macroeconomic risks in 2025?"
- "How did Munich Re's view on geopolitical risk change between 2024 and 2025?"
PDF Reports → pypdf extraction → page-tracked word chunking (800w, 100w
overlap; every chunk keeps the real page range its text came from, not
an estimate) → section detection → ChromaDB (cosine similarity)
↓
Hybrid retrieval: ChromaDB semantic search + BM25 keyword search, their
results unioned rather than one re-scoring the other, then a
cross-encoder (ms-marco-MiniLM-L-6-v2) re-ranks the combined candidate
pool on actual query/passage relevance
↓
LLM synthesis — cites (Company[, Year], p.N) inline; every
citation is regex-parsed back out and checked against the chunks the
model actually received before the answer is shown
Retrieval confidence bands (calibrated against this corpus: on-topic queries cluster ~0.55–0.60 cosine similarity, off-topic ones sit ~0.19–0.26) gate the whole pipeline — a query with no real answer in the corpus never reaches the model at all, and a weak-but-present match is shown with an explicit low-confidence banner instead of being presented as solid ground.
Why not send whole PDFs with the LLM's native citations API instead?
That approach (extracting citations directly from an attached PDF rather
than verifying a model-written page number) is genuinely more precise per
citation, and worth using — see pipeline/trim_pdf.py and the git history
for that direction. It requires manually trimming each report to its
narrative section first (the API caps PDF input at 600 pages, and most
annual reports run 300–600+), which doesn't scale to 15 companies × 2
fiscal years without a lot of by-hand work. This project optimizes for
broad, low-effort coverage instead: drop a PDF in data/raw/, run
pipeline/ingest.py, done — and closes most of the resulting gap with
real page-range tracking (not a linear estimate) plus ground-truth citation
verification against the retrieved excerpts.
Allianz · BASF · Bayer · Beiersdorf · BMW · DHL · Henkel · Infineon · Mercedes-Benz · Merck KGaA · Munich Re · SAP · Siemens · Siemens Energy · Volkswagen
git clone https://github.com/morichtereur/dax-intelligence
cd dax-intelligence
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
# Add PDF reports to data/raw/ named Company_Report_YYYY.pdf
python3 pipeline/ingest.py
streamlit run app/app.py
pip install -r requirements-dev.txt
pytest tests/ -v # chunking + page-offset unit tests, no API calls
.venv/bin/python eval/eval_retrieval.py # precision/recall of company retrieval against eval/gold_queries.json
.venv/bin/python eval/eval_grounding.py # citation grounding rate + a claim-level LLM faithfulness judge, real API calls
- Retrieval: ChromaDB (persistent, cosine similarity) + BM25 (
rank-bm25) + cross-encoder re-ranking (sentence-transformers) - Embeddings: ChromaDB default (all-MiniLM-L6-v2)
- LLM: Anthropic API
- Frontend: Streamlit
- PDF parsing/export: pypdf, reportlab
Word-based chunking (800 words, 100 overlap) over character-based — more stable across the mixed layouts of German annual reports. Each chunk tracks the real page range its words came from (pipeline/chunking.py), not a chunk-index/total-chunks linear estimate, which drifts further from the truth the longer and more unevenly-formatted the report. Section detection is heuristic (keyword matching on common annual report headers) rather than ML-based, which is fast and sufficient for this retrieval use case. ChromaDB chosen over FAISS for its persistent client and metadata filtering, useful for company- and year-level scoping without re-embedding.
anthropic
chromadb
streamlit
pypdf
python-dotenv
pycryptodome
rank-bm25
sentence-transformers
reportlab
watchdog
Built by Moritz Richter · Finance & Strategy Consultant · Zürich