Customer support
Answer frequent questions about products, policies, request status, and procedures, escalating sensitive cases to an operator.
AI services / Persian assistant
Persian chatbot and enterprise AI assistant
An assistant that answers only from your approved documents, cites its source, and hands off to a person when it does not know.
What it is for
A Persian enterprise assistant is designed for repetitive customer questions, helping internal staff find the right procedure, and drafting replies for human review. Direct actions on systems (placing orders, changing records) are added only after a successful pilot and a precisely defined authority boundary.
Common use cases
Internal knowledge assistant
Semantic search across regulations, contracts, procedures, and correspondence with the exact source clause shown.
Sales assistance
Draft replies to inquiries from the catalogue and approved terms, with final approval by the sales specialist.
Messaging channels
Deploy the same assistant on the website, Telegram, WhatsApp, or an internal messenger from one shared knowledge base.
Conversational intake
Collect initial request or lead details and record them in the existing system.
In-product assistant
Q&A over the data of a specific product or portal, such as the operations assistant in the Olbrich Portal.
Architecture and method
For enterprise Q&A our reference architecture is retrieval-augmented generation (RAG) with mandatory citations. The model is chosen after the data and constraints are understood, not before.
- 01
Approved knowledge base
Documents with an owner, validity date, and access level are chunked and indexed.
- 02
Persian retrieval
Normalisation of zero-width joiners, Arabic and Persian character variants, digits, and Jalali dates is applied before search.
- 03
Cited answers
Every answer points to a specific clause; without a source the assistant says it does not know and refers to a person.
- 04
Evaluation on real data
A versioned set of real user questions, reference answers, and supporting documents is built before deployment.
- 05
Deployment and monitoring
Conversation logging, specialist feedback, and periodic review of unsupported answers are part of operations.
Controls needed from day one
- Scope boundary
- Out-of-scope questions are defined and the assistant does not generate answers for them.
- Access control
- Each user only receives answers built from documents permitted for their role.
- Human hand-off
- The escalation path is defined with response hours and conversation context captured.
- Input security
- Prompt injection, data leakage, and tool misuse are reviewed against the OWASP guidance for LLM applications.
- Data hosting
- Deployment on the organisation's own infrastructure or hosting inside Iran is assessed against each customer's security requirements.
Zharf
Live demoTry the ZharfAI assistant right now
Ask about Persian assistants, document automation, or how to start a pilot.
Frequently asked questions about Persian chatbots
How is a rule-based chatbot different from an LLM assistant?
A rule-based bot only answers pre-written scenarios and stays silent otherwise. An LLM assistant understands varied questions but can confidently produce a wrong answer, which is why we recommend a retrieval-grounded architecture with mandatory citations and human hand-off.
Does the chatbot work with our existing software (CRM, ERP, portal)?
Integration can start with a controlled file, a work queue, or a software interface. The connection type is chosen after reviewing the system, access levels, and security requirements, and it is tested in the pilot.
How is Persian answer quality measured?
With a versioned acceptance set of real user questions. Four separate measures are reported: supported-answer rate, citation correctness, honest "I do not know" rate, and the severity of unsupported answers.
How much does an enterprise chatbot cost?
Cost depends on the question scope, document volume and quality, number of channels, integration type, and deployment model. After discovery you receive a staged estimate for a bounded pilot and then for expansion.
Is our data used to train models?
No. Your documents are used for retrieval and answering, and ownership of data and configuration stays with you. Data flow, hosting location, and retention are documented before the project starts.
Pick one channel and one question set for a pilot
In the first conversation we review the target channel, available documents, frequent questions, and acceptance criteria, including whether a pilot is appropriate at all.