Support fails the same way every time: the language model answers confidently but not by your rules. It does not know your prices, deadlines or exceptions, and it invents them fluently. The fix is not a bigger model but a chain: first find the right piece of your knowledge base, then hand it to the model as the source, then check the answer before it goes out. Each part covers a different risk and they are not interchangeable.
An embedding model turns the customer question and all of your internal documents into numeric fingerprints of meaning, and the system pulls the fragments that are closest in substance. Customers write in plain words while policies are written in formal ones, which is exactly where keyword search misses. Without this part the model answers from memory, meaning from an old snapshot of the internet rather than from your price list.
A model for meaning-based search in about a hundred languages. The core of RAG: the bot finds the right part of a document before answering.
Embeddings and rerankers based on Qwen3, among the best open ones for multilingual search, including Russian. VL versions search images, screenshots and video.
Proven models for semantic search. The multilingual versions work well with Russian and are still a reliable base for RAG.
A reranker takes the dozen fragments that were found and reorders them while reading the full question properly. It is cheap insurance: meaning-based search is fast but coarse, and the paragraph you need is often not the first one. Without it the first vaguely similar text goes into the answer, and the customer gets a rule from a different plan that does not apply to them.
Rerankers: they take passages found by search and reorder them by how well they actually match the question. v2-m3 is multilingual and lightweight, often paired with bge-m3.
Strong multilingual rerankers; m0 also ranks pages as images (scans, slides). The latest versions are open for non-commercial use only.
The language model receives the question together with the retrieved fragments and writes a human answer that points at its source. It is also asked to notice when a question falls outside the knowledge base and hand the chat to a person. The better the previous steps pick fragments, the easier this job is and the less you need the most expensive model available. Without this part customers read raw policy extracts.
The models that started mass open source in AI. A huge ecosystem of fine-tuned versions and tools.
OpenAI's first open models since GPT-2. Reasoning and tool calling; the smaller version fits on a single GPU.
One of the oldest Chinese open lines: from ChatGLM-6B to GLM-5.3. Strong at agentic tasks and programming; GLM-5.3-Flash understands images and is released under MIT.
A separate filter model looks at both the incoming message and the finished reply: is someone fishing for internal data, is another customer mentioned, is the bot promising something the company does not do. This filter is deliberately small and fast, it writes nothing and only raises a flag. Without it, one bad answer in a public channel costs more than the whole support saving.
Filter models that check chatbot requests and replies for dangerous topics against a list of categories. Version 4 also checks images. Russian is not officially supported.
Gemma-based filters: they check text for dangerous and offensive content, and ShieldGemma 2 checks images. Focused on English.
Safety filters for 119 languages, Russian among them. The Stream version checks a bot's reply while it is being generated and can cut it off on the fly.
The bottleneck is almost always the knowledge base, not the model. If your instructions contradict each other and were last updated long ago, the chain will reproduce that mess faithfully and faster. Before launch, collect fifty real questions from your inbox, run them through and read the answers yourself. Decide in advance which topics the bot must hand to a human without arguing: money, refunds, complaints and anything touching personal data, described here in general terms only, with a lawyer reviewing your case.
An open model can run on your own server: data stays in-house, there is no per-request fee, and the model can be fine-tuned on your documents.
