In this article
5 fictional documents · 14 indexed chunks
Chunks found
Answer
Pick a question or type your own.
What a good RAG does
- Cites document and section
- Ignores expired versions
- Says "I don't know" without a source
How it works inside
What collateral does a USD 80,000 SME loan need?
Example with fictional documents
- 1Current documentsOnly current versions are loaded: manuals, policies, contracts. Each with an owner and a date.
- 2Split into partschunkingEach document is cut into parts of one or two paragraphs. Each part remembers its document and section: that is where the citation comes from.
- 3Sort by meaningembeddingsEach part is placed on a map according to what it is about. Parts on the same topic end up together, even if they use different words.
- 4Find the closestsemantic searchThe question is placed on the same map. The system takes the closest parts: here, the collateral rules in the credit manual.
- 5Answer with the sourcelanguage modelThe AI writes using only those parts and shows the cited section. If nothing is close, it says it does not know and hands over to a person.
Example in a family company
It is not only for banks. In a distributor, the sales rep needs the current price list or the return policy while the customer waits.
Example: a sales rep checks the price list
Today
- 1Customer asksPrice and returns
- 2Rep callsThe office
- 3Someone checksWhich list is current
- 4~20 minAnswer (example)
With RAG
- 1Rep asksIn the internal chat
- 2SearchCurrent lists only
- 3AnswerCites "May list · §2"
- 4~1 minAnswer (example)
Does your case need it?
4 questions to decide
Question 1
Is the answer in internal documents: policies, manuals, contracts?
RAG, fine-tuning or prompt only
| Criterion | RecommendedRAG | Fine-tuning | Prompt only |
|---|---|---|---|
| What it does | Searches your documents and answers with them | Adjusts the model with your examples | You paste instructions and context by hand |
| When a document changes | Re-index that document | Retrain | Edit the prompt |
| Cites the source | Yes, document and section | No | Only what you pasted |
| Document volume | Thousands of pages | Not meant to store documents | Whatever fits in the context |
| Best for | Policies, contracts, manuals | Tone, format, repetitive tasks | Tests and small cases |
RecommendedRAG
What it does
- RAG
- Searches your documents and answers with them
- Fine-tuning
- Adjusts the model with your examples
- Prompt only
- You paste instructions and context by hand
When a document changes
- RAG
- Re-index that document
- Fine-tuning
- Retrain
- Prompt only
- Edit the prompt
Cites the source
- RAG
- Yes, document and section
- Fine-tuning
- No
- Prompt only
- Only what you pasted
Document volume
- RAG
- Thousands of pages
- Fine-tuning
- Not meant to store documents
- Prompt only
- Whatever fits in the context
Best for
- RAG
- Policies, contracts, manuals
- Fine-tuning
- Tone, format, repetitive tasks
- Prompt only
- Tests and small cases
More detail in RAG vs fine-tuning.
What it costs and when it pays back
The AI model is the cheapest part. What decides it is the preparation, the hosting and how much time your team saves.
What it costs and when it pays back
Monthly costUSD 152.05
Tokens: the unit AI providers bill for the text the model reads (input) and writes (output). OpenAI puts 100 tokens at about 75 English words.
Inside the model: Reading documents USD 1.20 (58 %) · Writing USD 0.75 (37 %) · Indexing USD 0.10 (5 %)
With a preparation cost of USD 8,000 to 20,000 (example range): 5.5 to 13.8 months.
Public prices per million tokens, OpenAI API pricing, checked 2026-10-11 (standard rate, short context): text-embedding-3-small USD 0.02; gpt-6-luna USD 0.10 / 0.50; gpt-5.6-luna USD 0.20 / 1.20; gpt-6.1-sol USD 2.00 / 10.00 (input / output).
Assumptions: 500 tokens per page (about 375 words; OpenAI estimates 100 tokens ≈ 75 words in English); per question, about 50 tokens to index it, 4,000 input and 500 output tokens. Preparation, hosting, minutes saved and cost per hour are example values: replace them with yours. For reference, Ecuador's 2026 basic wage is USD 482 per month, about USD 3 per hour: the real cost of a credit officer or a sales rep is higher.
Customer data in Paraguay and Ecuador
If the assistant will read complaints or customer files, review this with compliance before indexing. For banks and cooperatives: AI for banks and fintech in Paraguay and AI for banking and cooperatives in Ecuador.
Nov 2027
Paraguay, fully enforceable: November 2027. Personal Data Protection Law 7593/2025 takes effect 24 months after its publication on 2025-11-27 (art. 57). It includes rules on security and transfers.
2026-01-28
Ecuador: the SPDP issues its rule on domestic and international transfers of personal data (Resolution SPDP-SPD-2026-0004-R). Relevant if an AI provider processes data outside the country.
Third-party summaries: Ferrere on the Paraguayan law and Corral Rosales via vLex on the Ecuadorian resolution.
How to start
Before implementing RAG in your company
Example pilot, with deliveries every 2 weeks
- Week 1-2Documents and test questionsInventory, version clean-up and permissions.KPI: test set approved by the team
- Week 3-4First version with one teamIndexing, search and cited answers.KPI: % correct answers on the test set
- Week 5-6Real use and decisionTuning from feedback and a go/no-go to scale.KPI: time per query vs. today