Answers from your content
Your website, PDFs and documents form the knowledge base. Odeo answers from it, not from the model’s general knowledge.
Search thoroughly first, then answer correctly.
Advanced Hybrid RAG is Odeo's knowledge search. It understands follow-up questions, searches by meaning, keyword and spelling at the same time and merges the results. Only then does Odeo answer, with a link to the source and the date of the document.
- Understands follow-up questions
- Three search paths at once
- Latest version wins
Example follow-up: "And for businesses?" becomes the search query "Energy price 2026 for businesses". Odeo gives the business price with the 2026 tariff sheet as its source.
Example follow-up: "And for businesses?" becomes the search query "Energy price 2026 for businesses". Odeo gives the business price with the 2026 tariff sheet as its source.Example amount: the keyword search finds the line "washing machine CHF 150". Odeo names the CHF 150 subsidy with the 2026 subsidy programme as its source.Example typo: the spelling search recognises "electricty" as "electricity". Odeo points to the "Report an outage" page.Example latest version: the 2025 and 2026 tariff sheets list different base fees. Odeo takes the latest one and answers CHF 6.50 per month, as of 01.01.2026.What is RAG?
RAG stands for retrieval-augmented generation. Before the AI answers, it searches your own content for the right passages and hands them to the language model. The model then writes its answer from these passages, not from its general knowledge.
A language model knows a lot about the world, but nothing about your prices, your opening hours or your latest leaflet. RAG closes that gap without retraining the model.
The three letters
-
Retrieval
Search
Odeo searches your website, PDFs and documents for the sections that match the question.
-
Augmented
Add
The best sections go into the language model’s prompt, together with source and date.
-
Generation
Answer
The model writes the answer from exactly these sections.
What can an AI do with RAG?
RAG turns a language model into an assistant that knows your company. Your customers notice it with every question.
Names source and date
For every section, Odeo knows the link and the date. So it is clear where a statement comes from and how current it is.
-
Gives concrete values
Instead of just pointing to a page, Odeo gives the amount, the deadline or the opening hours from the document.
-
Says so when nothing fits
If the answer is not in your content, Odeo refers to your team instead of guessing.
-
Understands other languages
The search by meaning finds the right section even when the question is in French and the document is in German.
-
Stays current without training
New content can be found after the next import. The language model does not need to be retrained.
What happens between question and answer
The basic form, known as naive RAG, searches once by meaning and takes the closest hits. Odeo works in stages, and each stage solves a problem a basic search fails at.
When importing
Sections with context
When content is imported, every section gets a sentence that places it, for example "This section of the 2026 tariff sheet describes the energy prices for businesses". The document date and the language are added too. That way the search also finds sections that would be unclear on their own.
With every message
-
Stage 1: Understand follow-up questions
"And for businesses?" makes no sense on its own. Odeo rewrites the message into a complete search query based on the conversation so far, but only if it looks like a follow-up question. All other questions go into the search unchanged.
-
Stage 2: Three paths at once
The search runs by meaning, by keyword and by spelling in parallel. Meaning finds "costs for companies" in the section on business tariffs. Keywords find exact details such as "washing machine CHF 150". Spelling also finds "electricty".
-
Stage 3: Merge
The hits of the three paths are merged into one list. What several paths find moves up, duplicate texts drop out.
-
Stage 4: Re-rank7
A language model reads the best candidates with title, context and date and re-ranks them. So the section that really fits ends up on top.
-
Stage 5: Source and date
The best six sections go to the model with link and date. If two documents contradict each other, the model takes the latest version and names it.
-
Stage 6: Fast thanks to caching
The fixed part of the prompt comes first, the changing part last. That lets the fixed part be cached. The result: faster first words and lower costs.
-
Stage 7: Searches again by itself8
If the context provided is not enough, the model may search again by itself, with its own search terms and the same filters on your knowledge base.
Naive RAG or advanced hybrid RAG?
Naive RAG is the technical term for the basic form: one search by meaning, the closest sections go straight to the model. For many everyday questions that is not enough. Four examples:
- Follow-up
And for businesses?
Naive RAGSearches for "And for businesses?" literally and finds general texts about businesses.
Advanced Hybrid RAG Stage 1 Rewrites the question to "Energy price 2026 for businesses" and gives the price.
- Amount
Subsidy for a washing machine?
Naive RAGFinds texts about the subsidy programme, but not the line with the amount.
Advanced Hybrid RAG Stage 2 The keyword search finds "washing machine CHF 150", Odeo gives CHF 150.
- Typo
Report an electricty outage
Naive RAGA typo in a key term can make the hits worse.
Advanced Hybrid RAG Stage 2 The spelling search recognises "electricity" and finds the right page.
- Contradiction
How much is the base fee?
Naive RAGFinds the 2025 and 2026 tariff sheets and does not know which one applies.
Advanced Hybrid RAG Stage 5 Knows the date of every section and gives the 2026 amount.
| What the stages add | Naive RAG | Advanced Hybrid RAG |
|---|---|---|
| Search by meaning | ||
| Context for every section | ||
| Understand follow-up questions | ||
| Search by keyword and amount | ||
| Search by spelling | ||
| Remove duplicate hits | ||
| Re-rank hits7 | ||
| Source and date in the context | ||
| Latest version on contradictions | ||
| Prompt caching | ||
| Search again by itself8 |
Naive RAG and advanced RAG are stages from the technical literature on RAG. Naive RAG is the basic form: one search by meaning, the best hits go straight into the prompt. Advanced RAG adds steps before and after the search, such as rewriting the query and re-ranking.
What is the difference between RAG and an AI agent?
RAG makes sure the AI knows the right thing. An AI agent makes sure it does something: it decides on next steps by itself and uses tools to do so. Odeo combines both.
RAG: knowing
- Searches your content
- Answers with source and date
- Changes nothing in your systems
AI agent: acting
- Decides on next steps
- Uses tools, such as a search
- Reads and writes data in other systems
Odeo: both
-
Search as a tool
The model decides by itself whether to search your knowledge base again.
Go to stage 7 -
Workflows
Flows with fixed steps, their own model choice and hand-off to your team, from Professional.
Workflow builder -
Direct integration
Data from CMS, business application, CRM or ERP, without middleware such as n8n.
Integrations
Frequently asked questions
Answers on how it works, integration, data protection and pricing.
What is RAG?
RAG stands for retrieval-augmented generation. Before the AI answers, it searches your own content for matching sections and hands them to the language model. The model then answers from these sections and not from its general knowledge.
What is advanced hybrid RAG?
In the technical literature, advanced RAG is the stage above naive RAG, the basic form that searches once by meaning and passes on the best hits. Hybrid means that several search paths run at the same time. Odeo works hybrid and in stages: it understands follow-up questions, searches by meaning, keyword and spelling at the same time, merges the hits, re-ranks them and passes the best six sections to the model with source and date. If that is not enough, the model may search again by itself. Re-ranking is available from the Professional plan; the model’s own search is available when the chosen model supports tools.
What is the difference between RAG and an AI agent?
RAG provides the AI with the right knowledge from your content. An AI agent acts: it decides on next steps by itself, uses tools and works with other systems. Odeo combines both: the model may search your knowledge base again by itself, and with workflows and direct integration Odeo fetches data from your systems or hands requests over to your team.
What is the difference between RAG and fine-tuning?
Fine-tuning retrains a model with your data. That takes effort, and every change needs new training. With RAG the model stays unchanged, and Odeo searches your current content for every question. New or changed content can be found after the next import. Odeo relies on RAG, not on fine-tuning.
Does Odeo make up answers?
No. Odeo answers only from your knowledge base. If the search finds no matching section, Odeo says so openly and refers to personal contact instead of guessing.
What happens if two documents contradict each other?
Every section gets a date when it is imported, for example from the file name, the URL, the title or the PDF details, otherwise the import date. If two sources contradict each other, Odeo takes the latest version and names it in the answer.
Does Odeo find typos and exact amounts?
Yes. Besides the search by meaning, Odeo searches by keyword and by spelling. The keyword search finds exact details such as "washing machine CHF 150", the spelling search also finds misspelled words such as "electricty".
Does Odeo understand follow-up questions?
Yes. If a message looks like a follow-up question, such as "And for businesses?", Odeo rewrites it into a complete search query based on the conversation so far. Other messages go into the search unchanged, which saves time.
Do all these stages make the answer slower?
The stages are built for speed: the three search paths run at the same time, follow-up questions are only rewritten when needed, and re-ranking is skipped when there are only a few hits. The fixed part of the prompt is cached, which brings faster first words.
Is the content of different customers mixed?
No. Every search is strictly limited to the knowledge base of the customer in question. The search itself runs in our own database, without an additional search service.
What content can Odeo search?
Your website, which the crawler imports, as well as PDFs and documents you upload. Together they form the knowledge base, which you manage in the customer portal.
More features
See all features- Lead generation Spot, qualify and hand over leads to your CRM
- AI customer service Answers around the clock, hand-off to your team
- E-commerce AI agent Product advice, stock and order status in the chat
- Knowledge base chatbot Search documents, PDFs and policies by chat
- Booking assistant Take appointments, reservations and requests in a structured way
How thoroughly does your chatbot search today?
We show you how Odeo answers questions about your own content, with source and date.
Developed and operated by Sidora AG, Winterthur
Notes
- Re-ranking is available from the Professional plan; in the Starter plan, the order of the merged hits applies. It is skipped when there are only a few hits.
- The model's own search is available when the chosen language model supports tools.