LoginSign Up
← Back to all sparks
C

chattr

ANALYTICS
Velocity0.0

chattr — R package from the mlverse project

github.com/mlverse/chattr ↗

chattr deleted every LLM integration it had written and outsourced the lot to ellmer

llmrstudioide-integrationellmerarchitecture-change
◆Current state
chattr puts a large language model inside the RStudio IDE, either through a Shiny app or directly at the console. As of 0.3.0 it no longer talks to any model provider itself: all integration goes through ellmer, and the hand-written OpenAI, Databricks and LlamaGPT backends were removed. The package's supported model list is now whatever ellmer supports, and the Shiny app streams responses through ellmer rather than managing a background process.
◆Where it's heading
The first two releases show why that happened. Each provider brought its own error formats, token discovery and response handling, and 0.2.0 is largely a list of per-provider repairs — OpenAI error parsing, Copilot token discovery and model defaults, a new Databricks foundation model backend. Maintaining that surface scales linearly with the number of providers, and the pivot to ellmer trades it for a single dependency. The cost shows up immediately in 0.3.1, which exists solely to absorb a change in ellmer's token object.
◆Prediction
Expect chattr's releases to now track ellmer's, as 0.3.1 already does, with the package's own work concentrating on the IDE experience rather than model connectivity. New provider support will arrive without a chattr release at all.

◆Recent moves

  1. 0y ago

    Adapts to ellmer's token object change

    Adapts to a change in ellmer's token object. Three months after outsourcing all model integration, the first maintenance release is an upstream compatibility fix — the predictable ongoing cost of that decision.

    View source ↗
  2. 1y ago

    All model integration moves to ellmer, direct backends removed

    ⚡ SPARK

    Routes all LLM integration through ellmer and deletes the direct OpenAI, Databricks and LlamaGPT backends, so chattr now supports exactly what ellmer supports and nothing else. The Shiny app also drops its background-process streaming for ellmer's, which the entry describes as more error-prone. This ends the per-provider maintenance visible throughout 0.2.0.

    View source ↗
  3. 2y ago

    Databricks foundation models added; per-provider error handling fixed

    Adds a Databricks foundation model backend covering DBRX, Llama 3 70B and Mixtral 8x7B, and fixes how responses are routed once the current interface — console, app or notebook — is correctly identified. The remaining items are per-provider repairs: OpenAI error parsing and display, Copilot token discovery and a model default. Read against 0.3.0, this release is the argument for abandoning direct integrations.

    View source ↗
  4. 2y ago

    First release: LLM chat in the RStudio console and app

    The initial release, offering model interaction from a bundled Shiny app or the R console with backends for OpenAI, GitHub Copilot and LlamaGPT. All three of those direct integrations are removed two releases later.

    View source ↗