Candor Desktop IDE 2.0 grounds AI SQL generation in live database schema metadata

Candor has shipped version 2.0 of its Desktop IDE with a feature aimed at one of the most persistent failure modes of generalist AI assistants: SQL that looks plausible but references tables, columns or joins that do not exist. According to a company tutorial published on dev.to, the release grounds query generation in the live metadata of the connected database rather than leaving the model to guess from prior training data.

Grounding SQL in real metadata

The tutorial frames the problem bluntly: generalist assistants writing against complex warehouse schemas have no direct structural awareness, which routinely produces hallucinated table names, phantom column references and invalid joins that demand repeated manual debugging.

Candor 2.0 answers this with a metadata discovery step. Depending on the database dialect and connector, the tool reads catalog views and standard schemas such as information_schema.columns, and extracts:

• Table and column names
• Data types and nullability constraints
• Primary and foreign key relationships
• Dialect-specific syntax rules needed to build valid queries

The company is careful about the boundary here: only structural properties are pulled, never production rows, so customer and transactional data stays in the warehouse.

The workflow then runs in five stages. The developer connects to a warehouse or a local DDL catalog, Candor extracts the structural metadata, relevant table clusters are mapped into the model's context window using few-shot retrieval, SQL is generated from a natural-language intent, and the developer reviews the query, checks execution plans and runs it within their existing permissions.

What determinism buys in practice

The tutorial's worked example is a sales analytics query joining fact_orders and dim_customers to compute total gross revenue and distinct customer counts per acquisition channel. The interesting part is the join: the engine selects customer_key because foreign key metadata verifies it, instead of guessing the generic customer_id or id that a generalist model tends to reach for. Column selection works the same way, with gross_revenue_usd taken directly from the catalog rather than ambiguous candidates like amount or total.

There is also an economics angle. Rather than pasting an entire DDL that may run to megabytes into the prompt, targeted few-shot extraction limits context overhead, which the company says reduces API costs and latency.

The rest of the 2.0 release

Beyond schema-aware SQL, version 2.0 adds or extends:

• Frontier AI Chat, a conversational coding interface grounded natively in database and project schemas
• Bring Your Own Key support, routing requests through the developer's own LLM provider credentials or to local, air-gapped runtimes such as Ollama
• Built-in few-shot retrieval patterns intended to cut token consumption versus brute-force context dumping
• Extensions for Visual Studio Code, Cursor and Antigravity

Native binaries cover Windows and Linux; a macOS release and dedicated Claude extensions are described as in active development.

Availability and caveats

The IDE is distributed via candorsys.com, with a Community Edition offering a 14-day trial provisioned with 1 million tokens upfront. One caveat worth stating plainly: everything above, the determinism claims included, comes from the vendor's own tutorial rather than independent evaluation, so teams comparing it against generalist assistants will want to benchmark on their own schemas.

Why it matters

Text-to-SQL has long been reliable enough to demo and unreliable enough to distrust, and the gap usually has less to do with the model than with whether the model knows the actual schema. Candor's approach reflects a broader shift in the field: treat query generation as a metadata-grounding problem rather than a prompting problem, much as retrieval tamed factual hallucinations for documents. For data engineers, the specific failures being targeted — wrong join keys, deprecated staging views, invented columns — are exactly the ones that eat review cycles in production pipelines. The BYOK and local-runtime options matter too, since schema context is often the sensitive part that keeps warehouse tooling out of regulated environments. The grounding pattern will likely spread regardless of tool; whether Candor's execution of it holds up is what the trial period is for.