Curltale AI
Curltale AI
For decision makers
What it is, how it works, what teams use it for, how knowledge earns trust, and how to adopt it. The technical details are in the technical guide.
Overview
Most AI projects stall on the same thing: the assistant does not know how your organisation really works. Curltale AI gives it that knowledge, in two parts that are managed differently and connected on purpose.
What "active customer" or "open order" really means, how your systems connect, and what is sensitive. Agreed once, reviewed, versioned and published, then used for every answer.
Policies, procedures, product facts and the things experts carry in their heads. Captured from documents, web pages and interviews, checked against the source, and reviewed by people.
A policy can point at the business objects it governs. That link stays with the collection that holds the document, so the data model stays clean, and people who cannot read a collection never see the link.
Overview
Curltale AI sits between your systems and your AI. It does not replace your databases or your document stores. It gives them context.
Databases are read on request, read-only, and never copied into the platform.
Use the model provider you prefer, including models you host yourself, and a different one for different jobs if you like.
It runs in your environment. You decide where it lives and where documents are stored.
How it works
Two kinds of knowledge, built differently on purpose. One holds exactly what your data means. The other holds what your people know, and can prove it.
Your business objects, the exact definition of every word people use about them, how they join and what is sensitive. Agents do not write SQL from guesses: they ask for a plan, and the platform checks and rebuilds every statement under the caller's access before it runs, read-only.
Every document first becomes a source wiki: a clean page written from it and checked against it, line by line. Verified wikis are compiled into topics, one page per subject however many documents mention it. Claims keep the exact quote behind each fact, and relationships connect topics into a graph.
A topic is a complete article, so an assistant reads the whole argument: the price and its conditions, the rule and its exceptions.
Each claim records the passage of the verified page it came from and which version of that page. Anyone can click through to it.
A topic points at the business objects it is about, so the assistant knows which policy governs which part of your data.
How it works
The assistant takes the path the question needs. A number comes from your data with its definition attached. A rule comes from your documents with its source attached.
Before answering, the assistant checks what your words mean here: "active", a plan name or an abbreviation is looked up in the ontology, never assumed.
Every page is cut into chunks and indexed by meaning and by words, so an exact plan name or a paraphrase both find their way.
What comes back is the full topic page with the passage that matched and the topics linked to it: one call, the complete picture.
When nothing in the knowledge base covers a question, the assistant is told so and says so, instead of filling the gap from general knowledge.
How it works
Most AI knowledge tools index documents as they are and let a model sort it out at question time. Curltale AI does the careful work first, once, with people in the loop.
| The usual approach | Curltale AI |
|---|---|
| Documents are chopped into passages and searched as they are. | Each document is first rewritten as a verified page, checked claim by claim against the original, and anything doubtful waits for a person. |
| The assistant gets a handful of fragments and stitches an answer together. | The assistant gets whole topic pages by subject, plus the topics linked to them, already merged across every document that mentions the subject. |
| Sources are a list of file names at the end. | Every fact has a claim with its exact quote, the page version it came from and the passage it sits in. |
| When a document changes, old answers linger. | What a removed source supported is withdrawn, and the affected pages are revised surgically: only what nothing else supports is removed. |
| Contradictions are averaged away. | A change that contradicts a published fact is held, with both versions side by side, until someone decides. |
| Text-to-SQL guesses what "active" means. | The ontology says what it means, and every statement is checked against your access rules and rebuilt before it runs. |
| One big index for everyone. | Isolated collections with their own access and a public or internal tag that travels with every result. |
In practice
A customer-facing assistant answers from collections you mark public. Internal playbooks guide how it behaves without ever being quoted.
People ask in plain language. The agent uses your agreed definitions, so "active" means what the business means, and sensitive columns never leave.
Every claim carries its quote, its source and the page it came from. Contradictions between sources are held for a person to resolve.
An interviewer asks the questions a curious colleague would and turns the answers into reviewed pages, before the knowledge walks out the door.
New joiners and new AI tools get the same governed answers, and the same citations, from one place.
HR, finance and support each get their own collection with their own access. What one team adds never reaches another.
In practice
Responsibility sits with the people closest to the knowledge, and every change leaves a record.
| Who | Their part |
|---|---|
| The data team | Owns the meaning of the data. Drafts changes, reviews them and publishes, optionally with a second person's approval. |
| Knowledge editors | Add documents and interviews, and decide what the checker doubted. |
| Collection managers | Decide who can read or write a collection and whether its content may be shared with customers. |
| Administrators | Connect systems, choose AI models, manage people and access keys. |
| Everyone else | Ask, through the built-in assistant or their own AI tools, with exactly the access they have. |
In practice
Nothing reaches an assistant just because it was uploaded. Each document goes through five steps.
A claim is kept only if its quote appears word for word in the verified page. Anyone can click through from an answer to the exact passage.
When a source is updated or removed, what only it supported is withdrawn and the affected pages are revised surgically, checked against the sources that remain. Nothing quietly goes stale.
Control
The same rules apply whether a person or an AI agent is asking.
Each collection has its own list of who can read, write or manage it. Without access, a collection does not even appear to exist.
Every answer says whether it may be shared with customers. Assistants are told with each result, so they do not depend on someone remembering the rule.
Restricted values never leave your database. Personal ones can be masked. Every grant of access to sensitive data is logged.
Changes to the data model and to knowledge are reviewed by people before agents see them.
Every conversation, every lookup and every approval is recorded with who, when and what it cost.
A collection can be kept away from any AI model, or from all agents, whatever else is configured.
Getting started
Start with one domain and one assistant. The review queue and the audit trail grow with you.
Curltale AI starts small, as two services on your own server, and adds optional parts when volume grows. It is open source (MIT) and currently an open beta. Setup, the API and operations are in the technical guide, and the code and releases are on GitHub.
Getting started
The platform runs where you put it, and your databases are read, never copied. What reaches an AI model depends on what you switch on. You can keep any collection away from AI models, and you can use a model you host yourself.
Any provider that speaks the common OpenAI-style interface, hosted or self-hosted. Different jobs, such as writing, checking and interviewing, can use different models.
Yes. Agents start from the reviewed topic pages, which are complete, cited and linked to each other. For provenance they can see every claim with its exact quote, and for a detail a page does not carry they can search and read the verified page written from each document, or the original text. The same access rules apply at every step.
Access is per person and per agent, collection by collection, and for sensitive data, column by column. Administrators can see that a collection exists but must grant themselves access, which is logged, to read it.
Sources can be refreshed or replaced. Statements that lose their source are withdrawn, pages are revised, and changes that contradict what is already known wait for review.
A small team that owns the meaning of the data helps a lot. The knowledge side can start with just the people who own the documents.
Neither your databases nor your document stores. It sits between them and your AI, as the layer that gives them context and keeps the answers accountable.
Now, as an open beta. It is open source (MIT) and runs with Docker on your own server. The technical guide walks through setup, and releases are on GitHub.