Can AI learn from our past proposals and reports? Yes, but only if you choose how: retrieve, show examples, or retrain.
Short answer. Yes, but not automatically. A language model does not learn from your office’s work just by being used: its weights stay fixed. There are three deliberate ways to make it use your past proposals and reports: retrieve similar documents at the moment of writing, show it a few examples in the prompt, or fine-tune it on your documents. LoRA, a common fine-tuning method, trains a small set of added parameters and, in its authors’ GPT-3 tests, cut the number of trainable parameters by 10,000 times. For facts, one study found that retrieval consistently beat fine-tuning at adding knowledge. Past work also carries risks: client confidentiality, outdated prices and old mistakes. Curate what the system may learn from.
01
Three ways to make past work useful
Method
What the model sees
Updates when documents change
Best for
Risks
Retrieve
The closest past documents, fetched fresh for each task.
Immediately, once the index is refreshed.
Facts, clauses, prices and references.
Old or wrong documents surface. One client’s material can appear in another’s draft if the index is shared.
Examples in the prompt
A few chosen samples, pasted in with the task.
When you edit the prompt.
Structure and tone of a proposal or report.
The model copies details from the samples.
Fine-tune, for example with LoRA
Nothing extra at run time. The model itself has been adjusted.
Only after retraining.
House style, format and domain wording.
Needs data and expertise. Training data is baked into the adapter, so removing a document means retraining.
02
What the research says
Brown and colleagues showed in the GPT-3 paper that a model can perform a new task from examples given purely as text, with no gradient updates. That is the examples-in-the-prompt method. Hu and colleagues introduced LoRA, which freezes the pre-trained weights and adds small trainable matrices; against full fine-tuning of GPT-3 175B they report 10,000 times fewer trainable parameters and 3 times less GPU memory, with similar or better quality on the tasks they tested.
Ovadia and colleagues compared fine-tuning with retrieval for adding knowledge to a model. Retrieval consistently outperformed unsupervised fine-tuning, both for facts the model had met in training and for entirely new ones, and models struggled to learn new facts by fine-tuning. So use retrieval for what is true, and consider fine-tuning only for how to write.
03
What to curate before the system sees past work
Remove superseded versions, and mark old pricing and terms or exclude them.
Remove or mask client names, personal data and anything under a confidentiality clause that does not allow reuse.
Separate clients. OWASP lists cross-context leaks, where material from one group is retrieved for another, as a risk in shared indexes.
Pick the best examples, not all of them. Include a few winning proposals, and note why they won.
Record the date and the author of each document, so the system and the reader can judge age.
Check any document that contains an error you would not want repeated.
04
How an office would set this up
Begin with retrieval over 50 of your best proposals and reports, with metadata for client, date, topic and outcome.
Test with five real drafting tasks. Compare the output with what a person wrote and count facts that came from the wrong client or the wrong year.
Add prompt examples for tone, and write a one-page style guide.
Consider fine-tuning only after months of use, when you have hundreds of good examples and a way to test the result.
Measure time saved and errors found. Keep a person responsible for every document that leaves.
Review the curated set each quarter.
05
What ANVE does today
ANVE runs Private GPTs, assistants built on the office’s own material. Its documents function answers questions over the office’s own documents and shows the source passage, and its research maker, Scholar-Atelier, learns a writing style. Full document drafting with design templates is not on ANVE yet.
These run today in ANVE’s development build. No sealed unit exists yet.
06
Questions people ask
Does the AI learn from what our staff type into it?
Not by itself. A deployed model’s weights do not change as people use it. What changes is the documents you give it and the prompts you write. Cloud providers may use inputs for training under some terms, so check the contract; a local model does not send your inputs anywhere.
How many past documents does it need?
Retrieval works from the first document. Fine-tuning needs a set of good, consistent examples and a way to test the result.
Can it learn our writing style?
Yes, through examples in the prompt, retrieval of similar documents, or fine-tuning. ANVE’s research maker learns a writing style from past work.
Will it leak one client’s work into another client’s proposal?
It can if both sit in the same index with no separation. Keep clients apart, test for cross-client leaks, and have a person review each draft.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
