Reports always lag behind questions. While an analyst builds an extract the question has already moved on, and simple things like a weekly breakdown sit in a queue. This chain closes exactly that gap: a question in plain language becomes a database query, the result is explained in words, and a separate part checks that the answer actually follows from the data. That last part is what separates a working tool from a nice demo.
Before a question becomes a query, the system has to find the right tables and fields among hundreds. Meaning-based search does that: table descriptions and a business glossary become fingerprints, and each question gets the matching slice of the schema. Without this step the model either receives the whole schema, which is expensive and confusing, or a random slice and answers from the wrong tables.
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.
Alibaba embeddings and rerankers for search: from tiny to 7B based on Qwen2. There is a multilingual mGTE version with long context.
A specialised model writes a valid database query from the wording of the question and the schema description. Narrow models built for this are usually more accurate and cheaper than general ones, because queries are exactly what they were trained on. Access must be read-only and role-based: the chain should be physically unable to change anything. Without this part everything queues behind an analyst.
One of the first open models that turn a plain-language question into an SQL query against a database. Available in Ollama, but newer competitors are already stronger.
Alibaba models for turning questions into SQL, based on Qwen2.5-Coder. They work with different SQL dialects; a small 3B version suits modest hardware.
Models for turning questions into SQL, trained on millions of synthetic query examples across different databases. Three sizes for different hardware.
A language model takes the returned table and turns it into a readable answer: what grew, how it compares with the previous period, what the data does not cover. It also asks back when a question has more than one reading, for example whether revenue should include tax. Without this part a manager gets a table and still walks over to an analyst for a summary.
A family of language models with strong Russian language support, from small versions for a laptop to a flagship on par with commercial APIs.
DeepSeek's flagship line: from the first 7B/67B to V4-Pro with 1.6 trillion parameters. Closed-model quality under an open MIT license; V4-Flash-Vision-Exp and V4.1-Flash understand images, context up to 1M tokens.
OpenAI's first open models since GPT-2. Reasoning and tool calling; the smaller version fits on a single GPU.
A separate checking model compares the written answer against what the database actually returned and flags claims the data does not support. That is cheaper than working out afterwards why a wrong number was quoted in a meeting. Without it, mistakes are visible only to someone who knows the data by heart, and that person is usually not the one reading this report.
An open judge model: it scores other models' answers against your criteria and explains the score. A replacement for paid models in the reviewer role.
Checks whether each claim in an AI answer is supported by the source documents. The small versions are free; the larger 7B is in Ollama but non-commercial.
A small model that checks whether an AI answer is grounded in the source text or made up. Runs on a CPU and works well as a filter in RAG systems.
The chain works only as well as your database is understandable. If fields are named in abbreviations and the same metric is computed three different ways, the model inherits that confusion. Start with one data mart and a dozen agreed metrics with written definitions rather than the whole warehouse. Read-only access, role separation and a query log are mandatory, otherwise a convenient tool becomes a leak.
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.
