Chatbot and RAG
ALGOFACT builds chatbots and RAG (retrieval-augmented generation) assistants that answer questions from internal knowledge such as procedures, contracts, technical documentation and support tickets, with a citation for every answer. Each user only gets documents they are already allowed to see, the assistant links to the passage behind its answer and it says so when it cannot find a source. It works in English and Turkish and runs in the cloud, on premises or hybrid, with GDPR and KVKK requirements handled at design time.
Last updated:
If employees or customers keep asking the same questions in different words, and the answer is written down somewhere but takes too long to find, a RAG-based assistant is the right place to start. The assistant takes the question, retrieves the relevant document passages, hands only those passages to a large language model and returns the answer with its source. The model is not retrained on your documents, so when a procedure changes, the answer changes with it.
If your documents are scattered, contradictory or out of date, the first job is content, not an assistant. No retrieval method turns a wrong procedure into a right one. In the discovery call we look at the state of your content together and, where needed, point out which documents should be cleaned up before the assistant goes live.
These three get mixed up often. What the user expects at the end decides which one you need.
Value shows up fastest where the answer already exists in writing but searching for it eats time.
Retrieval is filtered by the identity of the person asking. Someone who cannot see a document in the source system cannot see it through the assistant either, and permissions are re-read from the source on every indexing run.
Every answer includes the name, section and link of the document it relies on. The question, the retrieved passages and the generated answer are logged together for audit.
If the retrieved passages do not cover the question, the assistant does not guess. It says it could not find the information and suggests the team to contact, and this behavior is measured on the question set.
Text written inside a document cannot change the assistant’s rules. Every uploaded file is treated as a knowledge source, never as a command.
If documents contain personal data, the requirements of the GDPR or Türkiye’s Law No. 6698 are handled in the indexing design: unneeded personal fields stay out of the index and logs are masked. If data must not leave your environment, the language model and the vector index run on your infrastructure. Legal interpretation stays with your counsel.
The assistant is built to answer only from retrieved sources and to cite them every time. Answers users flag as wrong are reviewed, the cause is found (a missing document, poor splitting, a retrieval miss) and the question is added to the evaluation set.
Yes. Turkish is an agglutinative language, so keyword search alone falls short. We combine multilingual embedding models for semantic search with stem-based keyword search and measure which combination works best on your documents with the question set. A question in English can be answered from a Turkish source, with the citation pointing to the original.
Most text-bearing sources: PDF, Word, Excel, PowerPoint, HTML, wiki pages and support tickets. Scanned documents are converted with OCR; accuracy on handwriting and low-quality scans is checked separately during the pilot.
If a commercial model API is used, we select enterprise terms under which your data is not used for training and retention is acceptable. If those terms are not enough, we run an open model on your own infrastructure. The decision is made together, based on your privacy requirements and the question set results.
The number of source systems, document volume and quality, the share of scanned documents, how complex the permission model is, the number of channels and where the model runs. We present the scope and the expected running cost in writing after the discovery call.
Describe the problem briefly and we will reply with a time for a 30-minute discovery call.
Book a discovery call