The fastest-moving corner of applied AI, explained simply. What RAG is, why the classic version is already fading, and five papers, one map and four applied, that change how you would actually build a financial AI system.
A warning before the list: this is the fastest-moving corner of AI, and it shows. Some of the ideas below are already ageing at the edges, and at least one of the papers says exactly that about itself. Read them for the durable ideas, not the leaderboard numbers, which turn over month to month.
The industry has not settled its own vocabulary, so start here. A single-turn LLM call is a one-shot: prompt in, answer out, like a chatbot or a sentiment classifier. A workflow chains several LLM calls along a predefined path, each step finishing before the next begins. An agent is handed a goal and decides its own steps, calling tools, checking results and correcting course. Autonomy rises across the three, and the CFA Institute's exhibit further down draws exactly that line.
RAG means retrieval-augmented generation. Instead of trusting whatever a model memorised in training, you first retrieve relevant documents and place them in the prompt, so the answer is grounded in real, current sources rather than recall. The classic recipe: chop documents into chunks, embed each as a vector, and at query time pull the nearest chunks by similarity.
That recipe is quietly losing ground. Frontier models now carry context windows of hundreds of thousands to millions of tokens, so for many tasks you can skip the brittle chunk-and-embed pipeline and simply hand the model the whole document. Where retrieval still matters, at real corpus scale, the winning designs are agentic: the system retrieves, judges whether the evidence is actually enough, reformulates and retries, routes on metadata, and reranks. The four applied papers below are that shift, in order.
Here is my hands-on version of these ideas, a three-agent system that turns out dozens of DCF models in Excel a day:
If you only read one, start with paper 02, the reality check the rest of the hype needs. Then use the retrieval and multi-agent papers as a build blueprint.
Before the applied work, a map helps. The CFA Institute's report on agentic AI for finance lays out where these systems plausibly fit across the investment process, and, usefully, draws the line between single-turn calls, workflows and full agents on one axis of autonomy. Its exhibit is the cleanest picture of that distinction, reproduced below.
Treat it as orientation rather than gospel. An industry-wide survey of a field this young dates in parts even as it publishes, so read it for the taxonomy and the shared vocabulary, then ground each claim in the applied papers that follow.
Read it for the map. The field moves fast, so parts already date, mix it with the applications below.
This is the paper to read first, because it inoculates you against the rest of the hype. It reviews 77 studies of LLM and agent-based trading systems and finds the field failing its own basic tests: barely any of them model real trading costs, and none are reproducible.
The implication is blunt. Most headline "the AI beat the market" results quietly evaporate once you add slippage, fees and an honest out-of-sample split. Read it as a checklist for what a credible evaluation must include, so you can tell a real result from a backtest that only looks like one.
77 studies, barely any model real costs, none reproducible, your skeptic's shield, read it first.
Most RAG systems reach for a vector database by default. This paper runs the head-to-head that default rarely gets: keyword search (BM25), dense vector search, hybrid, and corrective RAG with a reranking step. The result is consistent, hybrid retrieval plus a reranker wins, and pure vector search, the industry default, is frequently the wrong choice.
There is a finance-specific lesson buried in it too. Tables, the heart of filings and factsheets, need their own handling; feed them through naive text chunking and you shred exactly the numbers you were trying to retrieve.
Hybrid plus reranking wins, pure vector is the wrong default, and tables need their own handling.
Classic RAG retrieves once and answers. This paper closes the loop: the agent retrieves, judges whether the evidence is actually sufficient, and retries with a refined query when it is not. Only when the evidence clears the bar does it answer.
That read-and-judge loop is the current frontier pattern for document QA, and it is a perfect fit for finance. It spends a few extra calls to buy far fewer confident-but-wrong answers, which is exactly the trade you want when the source is a prospectus and the reader is going to act on the number.
Retrieve, judge, retry, the read-and-judge loop trades a few calls for far fewer wrong answers.
MimirRAG is the one that reads like a blueprint you can copy. It is a multi-agent system that filters on metadata before semantics: narrow to the right issuer, document, period and section first, then run semantic search inside that thin slice rather than across the whole corpus.
For finance that ordering is the whole ballgame. The same phrase means different things across issuers, filings and years, and semantics-first retrieval happily returns the right words from the wrong document. Route on metadata first, then rerank (paper 03) inside the result, and you have the most copyable design here.
Metadata before semantics, the most copyable blueprint for financial RAG.
Put together, these five give you a working stance on agentic AI in finance: keep a healthy distrust of unaudited results (paper 02), retrieve with hybrid search and reranking rather than pure vectors (paper 03), close the loop so the system judges its own evidence (paper 04), and route on metadata before semantics in a multi-agent design (paper 05), with the CFA map (paper 01) to place it all. The video above is my hands-on version of the same ideas.