Embedding Model Fine-Tuning
For RAG and search systems that keep retrieving the wrong passages. We measure retrieval on your own queries, try the cheaper fixes first, and fine-tune the embedding model only when they fall short, then prove the gain before you re-index.
Bring 20 questions your search answers badly. We'll tell you whether it's a chunking, search, reranking or embedding problem.
You get retrieval scores on your own queries before any training starts.
We build and hand over the model. Running it in production is a separate engagement.
What is embedding model fine-tuning?
An embedding model turns text into lists of numbers so a search system can find passages that mean the same thing as a question. General embedding models are trained on public text, so they can miss what matters in yours: product codes, internal jargon, legal or medical terms, the way your users phrase questions. Embedding model fine-tuning trains the model further on pairs of real questions and the passages that answer them, taken from your own documents, so that it places them closer together. When retrieval is the weak link in your RAG system, better retrieval means the language model gets the right material to answer from. It's one of three methods in our custom AI model development work.
Signs retrieval is your problem
- The answer is in your documents, but the assistant says it can't find it, or answers from the wrong one.
- Search misses exact terms: part numbers, case references, product names.
- Results are close but not right: the right topic, the wrong policy version or product line.
- Switching to a bigger general embedding model didn't help.
If the right passage is retrieved and the answer is still wrong, the problem is the language model or the prompt, not retrieval. We check which in the first step.
Embedding model or reranker?
A search system usually has two stages. The embedding model finds a shortlist of candidate passages quickly from the whole collection. A reranker then reads each candidate with the question and puts the best ones first. If the right passage never makes the shortlist, a reranker can't save it, and the embedding model is the problem. If it's in the shortlist but ranked low, a reranker may be enough. We measure both before deciding.
Cheapest fix first
| Step | What it fixes | What the evidence says |
|---|---|---|
| 1. Measure retrieval on your queries | Shows whether retrieval is the bottleneck at all | Databricks: fine-tuning “works best when retrieval is a clear bottleneck” (2025) |
| 2. Check chunking (how documents are split into searchable pieces) | Chunks that are too small, too large or split mid-thought | Weaviate lists it as a check before fine-tuning |
| 3. Add keyword search (hybrid search) | Exact codes, IDs and rare terms | Weaviate recommends trying it first for exact-match queries; Cisco IT still needed keyword search after fine-tuning |
| 4. Try a domain-specific embedding model | Legal, financial, medical or code vocabulary | Off-the-shelf domain models exist for some fields. We test them against your queries |
| 5. Add an off-the-shelf reranker | The right passage is shortlisted but ranked low | Databricks: reranking helped, but “marginal on some datasets” |
| 6. Fine-tune the embedding model | Your vocabulary and question style, where steps 2–5 fall short | Databricks: fine-tuning improved retrieval on two of three enterprise datasets and beat reranking on two |
| 7. Fine-tune or distil the reranker | Ranking quality, or reranker speed | Less published evidence. Measured case by case |
Many teams stop at step 3 or 5, and the fit memo tells you if you're one of them. Building the RAG system itself is RAG Development; this page is about the models inside it.
Not sure where your system is failing?
Book a free fit call(opens scheduling widget)How we fine-tune an embedding model
- 01Build a query set. 50–200 real questions with the passages that answer them, checked by your domain expert. This is the test set, and training never sees it.
- 02Create training pairs. Usually synthetic: a language model writes questions that each passage in your documents answers. Databricks fine-tuned “on synthetic data… without manual labeling”. Your expert reviews a sample.
- 03Pick base models. Two or three candidates (general and domain-specific, open-weight or managed), chosen for language, size and licence.
- 04Train and score. Fine-tune each candidate and score it on the query set with Recall@10 (how often the right passage is in the top 10) and nDCG (whether the best passages come first).
- 05Check the answers. Run a sample of questions end to end through your RAG system, old vs new, so you see the effect on what users read.
Re-indexing and switching over
A new embedding model means every document has to be converted again, and the index rebuilt, because old and new numbers can't be compared. That's real compute and time, so the fit memo estimates it before you commit. We hand over a re-index plan, run the new index alongside the old one, switch once the query set confirms the gain, and keep the old index for a quick rollback.
Which embedding models can be fine-tuned
- Open-weight embedding models (many on the public MTEB benchmark) can be fine-tuned and run in your own environment. Most of our work uses these.
- Google supports tuning for text-embedding-004, text-embedding-005 and text-multilingual-embedding-002, and reports gains of “up to 41% (average 12%)” on public retrieval benchmarks (vendor-reported).
- AWS Bedrock lists Amazon Titan Multimodal Embeddings G1 for fine-tuning.
- OpenAI's embeddings documentation doesn't offer fine-tuning. For other hosted providers, check whether they offer it.
(Checked November 2026.) A model tuned on a managed platform runs there; the fit memo compares both routes.
What changes when it works
- The right passages come back, measured as Recall@10 and nDCG on your own queries, before and after.
- Better answers, checked end to end on a sample of real questions.
- Your exact terms are found, through hybrid search, the tuned model, or both.
- A safe switch, with the old index kept until the new one is proven.
- A query set you keep, to test any future embedding model on the same questions.
We don't quote retrieval gains before measuring your current system. Databricks' own test showed fine-tuning doesn't always win, and we'll tell you if it won't for you.
How an engagement runs
Fit (measure retrieval; check chunking and hybrid search; go/no-go at Gate 1) → Baseline (query set approved by your expert) → Data (synthetic training pairs, expert-checked sample) → Train (two or three candidates) → Prove (retrieval scores, end-to-end answer check, latency; Gate 2) → Hand over (model, re-index plan, runbook; your team re-runs the pipeline once).
If the Fit Sprint finds that chunking, hybrid search or a reranker fixes the problem, it ends with that recommendation, and RAG Development can implement it.
How you buy it: free fit call → Fit Sprint → Build. The Fit Sprint and the Build are each priced per scope.
What you receive
Fit memo (where retrieval fails and the cheapest fix) · query set with relevance labels and baseline scores · training pairs and pipeline · fine-tuned embedding model (and reranker, if in scope) with run log · evaluation report (retrieval scores, answer check, latency, failure cases) · re-index plan, deployment package and runbook.
You own the trained model, the dataset and the evaluation set once the work is paid for; we keep our own reusable tooling, and the base model's licence still applies. Need it connected to your product? We can scope that integration into the Build.
What we need from you
A sample of your documents (or index access) · 20–50 real user questions to start · a domain expert to judge which passages answer them · details of your current retrieval setup (embedding model, chunking, vector database, any reranker) · the right to use the documents for training · a cloud account for compute, including re-embedding.
Not included
Building or rebuilding the RAG application or vector database (RAG Development) · running the re-index in production (we plan it; your team runs it) · hosting the model · fixing answer quality when retrieval is already right (that's the language model; see AI model fine-tuning).
How to judge us
| Question | Crescent AI | Vector database or platform tutorial | Doing it in-house |
|---|---|---|---|
| Finds out whether retrieval is the problem | Yes, step 1 | Assumes it is | If you build the query set |
| Tries cheaper fixes first | Yes, in a written order | Sometimes mentioned | Your call |
| End-to-end answer check, not just retrieval scores | Yes | Rarely | Your call |
| Re-index and rollback plan | Yes | Rarely | Your team |
| Best when | You want the diagnosis and the fix from one team | You have an engineer free and a clear domain gap | You have ML engineers and time |
Who leads this work
Yash Amin, founder of Crescent AI. About Crescent AI
Common questions
Should I fine-tune my embedding model or use a reranker?
Measure first. If the right passage isn't in the shortlist, fix the embedding model or search. If it's shortlisted but ranked low, try a reranker.
Which embedding model should I use?
The one that scores best on your own questions, which is often not the top of a public leaderboard. We test two or three, including domain-specific models where they exist.
How much data do we need?
For testing, 50–200 real questions with known answers. For training, enough question–passage pairs to cover your topics, which we usually generate synthetically from your documents, so manual labelling stays small.
Can we fine-tune OpenAI's embedding models?
OpenAI's embeddings documentation doesn't offer it. Google and AWS Bedrock offer tuning for some of their embedding models, and open-weight models can always be fine-tuned.
Do we have to re-index everything?
Yes, if you change the embedding model. We estimate the cost up front and keep the old index until the new one is proven.
Is there an embedding model for legal or financial documents?
Some domain-specific models exist. We test them against your questions alongside general models before deciding to train.
Bring us 20 questions your search answers badly.
We'll tell you where retrieval is failing and the cheapest way to fix it, before anything is trained.
Mutual NDA before you share data · Fixed-scope Fit Sprint · Retrieval scored before any training
Part of Custom AI Model Development · RAG Development · AI Model Fine-Tuning · Model Distillation