Parse what others skip
Item hierarchy, cross-references, footnotes and XBRL tags, plus the tables that hold most of the numbers and defeat naive text chunking.
ICAIF 2026 · Hands-on tutorial · Milan
Trustworthy Document Intelligence in Finance.
Build a question-answering system over real filings and earnings releases, then prove that every answer, and every number in it, traces back to its source.
Organized with FinHorizons Foundation · Listed on ICAIF 2026 tutorials
Why this tutorial
Every answer traced to its source.
Question answering over filings, earnings calls and research documents is the most widely deployed LLM use case in financial institutions, and the one where hallucination is least tolerable. Generic RAG tutorials skip what makes financial documents hard. This one does not.
Trace every number
Hybrid retrieval with structured citations, and numeric grounding that links each figure in an answer to a source cell or sentence.
Measure, then fix
Claim-level verification and citation-accuracy metrics on an evaluation set built from your own corpus. "Sounds plausible" is not a score.
What you take home
A runnable pipeline
From raw filing to grounded answer with citations, open source and yours to keep.
A faithfulness harness
Claim extraction, source verification and citation scoring, reusable on your own corpora.
Design guidance
The choices that matter most: parser, chunking, retriever, re-ranker and grounding checks.
Prerequisites Python proficiency and basic familiarity with LLMs. No document-AI experience is assumed. Open-source models and tooling throughout, with instructions for substituting commercial APIs.
Curriculum
Four blocks. One pipeline.
From raw filing to verified answer, with a checkpoint after each block and a hands-on exercise in Colab.
-
Parsing and indexing
Structure that survives extraction.
- Structure-aware parsing of filings: items, footnotes and XBRL tags
- Table and figure extraction with multimodal models
- Chunking strategies that respect document structure, compared
Exercise Parse a 10-K and recover its financial-statement tables intact.
-
Retrieval and generation
From question to cited answer.
- Hybrid dense and sparse retrieval with re-ranking
- Generation with structured citations
- Numeric grounding: every figure traced to a source cell or sentence
Exercise Build the question-answering system and answer analyst-style questions with citations.
-
Faithfulness evaluation
Measured against the source.
- Building an evaluation set from your own corpus
- Claim-level verification and citation-accuracy metrics
- Hallucination detection and stress cases
Exercise Score your system, find its worst failure, and fix it.
-
Production concerns
What changes after the demo.
- Point-in-time corpora across restatements and amendments
- Incremental indexing, cost and latency
- Audit trails and open research problems
Discussion The design choices that matter most: parser, chunking, retriever, re-ranker and grounding checks.
Program
Two hours, hands-on.
Timings · CET
Saturday, 14 November 2026, 08:30–10:30 CET, Room 3 at Bocconi University. The notebook runs in Colab over a prepared corpus and replays a recorded run, so a laptop and a browser are all you need: no GPU.
Part 1 · Build the pipeline · 08:30–09:35
08:30–08:40Opening
The financial document stack, why filings break generic RAG, and the anatomy of the tutorial pipeline.
Yaxuan Kong08:40–09:05Parsing and indexing
Structure-aware parsing of filings, table and figure extraction with multimodal models, and chunking strategies compared. Exercise: parse a 10-K and recover its financial-statement tables intact.
Yaxuan Kong09:05–09:35Retrieval and generation
Hybrid retrieval, re-ranking and structured citations, with numeric grounding that traces every figure to a source cell or sentence. Exercise: build the question-answering system and answer analyst-style questions with citations.
Yaxuan KongPart 2 · Evaluate and deploy · 09:40–10:30
09:40–10:05Faithfulness evaluation
Building an evaluation set from your corpus, claim-level verification, citation-accuracy metrics and stress cases. Exercise: score your system, find its worst failure, and fix it.
Yaxuan Kong · Yichen Li10:05–10:20Production concerns
Point-in-time corpora, incremental indexing, cost and latency, audit trails, and open research problems.
Yichen Li10:20–10:30Wrap-up and Q&A
Closing discussion, including the questions submitted in advance through this site.
Yaxuan Kong · Yichen LiTimes are local to Milan (CET). At 09:35, between the two parts, five minutes of live questions: anyone in the room asks the system just built, from a phone, at docai.finhorizons.org/ask. Each block ends with an exercise of about ten minutes and a checkpoint cell that prints the expected result. The presenters circulate during the exercises.
Materials
Open source. Yours to keep.
The slides and the notebook are in one public repository, ICAIF 2026 Tutorial Materials.
The checklist, the resource hub and the live demo are on this site.
-
Slides
One HTML file you can click through, with the exercises inside, plus PDF and PowerPoint copies.
Slides on GitHub → -
Notebook
The whole pipeline in one Colab notebook. No GPU needed: it replays our recorded run.
Open in Colab → -
Evaluation checklist
250 checks for evaluating question answering over financial filings.
Open the checklist → -
Resource hub
More than 800 papers, tools and data sets, each placed at its step of the pipeline.
Open the resource hub → -
Ask the documents
Ask NVIDIA's filings anything. Every number in the answer is checked against its source.
Ask a question → -
Filings and test questions
NVIDIA's four filings, with fifteen test questions and their checked answers.
Data on GitHub →
Code under the MIT license, text and slides under CC BY 4.0. The SEC filings are public records. github.com/finhorizons/ICAIF_2026_Tutorial_Materials
Questions in advance
Send your questions before the session.
Working on document question-answering, or planning to? Tell us what you would like the tutorial to address. We read every question while preparing and build the common ones into the session and the closing Q&A.
- Anything from a specific block to your own documents and use cases.
- Short questions are welcome. No need to attend to ask.
- Prefer email? Write to events@finhorizons.org.
People
The people behind the tutorial.
Organized by Yaxuan Kong, Yichen Li and Stefan Zohren. Yaxuan and Yichen present.
Yaxuan Kong
University of Oxford · FinHorizons Foundation
Organizer · PresenterYaxuan is a doctoral researcher in Engineering Science at Oxford, working on the reliability and trustworthiness of LLM-based systems in financial decision-making, bias auditing for financial LLM applications, multimodal financial reasoning, and financial time series. She has published at ICML, ACL, AAAI and ICAIF and in The Journal of Portfolio Management, and brings practical experience from an AI-driven hedge fund and Citigroup. She co-created and co-taught a hands-on summer school tutorial at Oxford in June 2026.
Yichen Li
Organizer · PresenterYichen is a founding member of the FinHorizons Foundation, where he leads its work on AI in financial services, from investment and asset management applications to the trustworthy and responsible deployment of AI systems, including fairness auditing and governance. He holds a PhD in Finance from Durham University and has worked as an investor and finance entrepreneur, with extensive experience of deploying AI in real-world financial settings and of translating research into practice for institutional audiences.
Stefan Zohren
University of Oxford
OrganizerStefan is Professor of Machine Learning in Finance at Oxford, former Deputy Director of the Oxford-Man Institute of Quantitative Finance, and Principal Quant and Scientific Advisor with Man Group. His research covers applied machine learning in finance, including deep learning, NLP approaches and generative models. He is a Turing Fellow and received the 2023 Bernstein Fabozzi/Jacobs Levy Award. He has many years of experience teaching machine learning in finance to graduate students and practitioners.