Every technical conversation I have eventually arrives at the same question: should we fine-tune a model on our documents, or retrieve from them at query time? I've built both into StockLens, for different pieces of the product, and the honest answer is that the question itself is usually a little bit wrong. It's not either/or. But if you forced me to pick a default, retrieval wins for almost everything a research desk actually needs.
Why retrieval wins for filings and transcripts
Financial documents have a property that breaks the case for fine-tuning: they're current by definition. A 10-Q filed this morning is only useful because it's new. A model fine-tuned on last quarter's filings would be confidently, fluently wrong about anything that changed, and it wouldn't tell you it was wrong. That's the failure mode that worries me most in this domain: not the model refusing to answer, but answering smoothly from stale memory.
Retrieval-augmented generation sidesteps this by design. The transcript intelligence layer in StockLens doesn't "know" what a company said on its last call from training, it looks it up, every time, from the actual document, and returns the answer with a citation back to the page. When a document changes, the index updates and the system is current within minutes, not whenever the next fine-tuning run happens to be scheduled.
There's a second reason retrieval wins that's less discussed: it's the only approach that gives you a citation for free. Fine-tuned models bake facts into weights; there's no clean way to point back at "this came from page 14, paragraph 3." For a research desk that has to justify a thesis to an investment committee, the citation isn't a nice-to-have, it's the entire point.
Where fine-tuning actually earns its cost
None of this means fine-tuning is useless, it means it's aimed at a narrower target than most teams initially assume. The place it earns its keep is form, not facts: teaching a model the specific vocabulary, structure, or tone that generic retrieval-plus-reasoning keeps fumbling.
Sentiment classification on earnings call language is a good example. "We're cautiously optimistic about the back half" reads very differently depending on which management team said it and how they usually communicate. A lightweight model fine-tuned specifically on labeled financial language consistently outperforms a general-purpose model asked to classify tone from scratch. It's not answering a factual question that changes quarter to quarter, it's recognizing a pattern that stays relatively stable, which is exactly what fine-tuning is good at.
The other place I'd reach for fine-tuning is house style: if a fund wants every generated research note to match a specific format and level of formality, a lightly tuned model gets you there faster and more consistently than prompting alone.
The hybrid pattern I actually deploy
In practice, the systems I build almost always end up hybrid, and the split is consistent: retrieval for facts, a tuned or specialized model for form and classification, and a general reasoning model on top deciding how to combine the two. Concretely, in the transcript intelligence pipeline: documents get parsed and chunked, relevant passages get retrieved against the question being asked, a sentiment-tuned pass runs across the retrieved language for tone, and a reasoning model synthesizes the final answer with citations attached to every factual claim.
No single piece of that pipeline is doing something exotic. What matters is the discipline of keeping facts and form separate, and never letting a model's fluency substitute for an actual source document. That discipline is more important than the choice of model, and it's the thing I'd tell any team building in this space to get right before anything else.
The question to actually ask
So the useful question isn't "RAG or fine-tuning." It's: for this specific output, is the value in being current and citable, or in matching a stable pattern reliably? Filings, guidance, news, anything with a timestamp that matters: retrieval. Tone and classification of language that doesn't change its meaning quarter to quarter: fine-tuning has a real case. Most financial research workflows need both, in that order.