rtemis.llm NEWS

0.8.8

• Compatible with S7 releases after 0.2.2: the namespace imports data.table’s := and excludes S7’s, which removes the import-replacement warning on load, and load, unload, and build hooks use S7_on_load(), S7_on_unload(), and S7_on_build() when S7 provides them, with methods_register() for earlier S7 releases. The minimum S7 version is unchanged.
• Added concurrency to llmapply(), dmapply(), agentapply(), and map(), with independent agent state per input, bounded local workers, and batch_elapsed timing.
• Anthropic requests follow the rules of the configured Claude model, including Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5: sampling parameters, thinking, effort, structured output and tool choice are sent in the form the model accepts, and a value the model rejects stops before the request. Structured output uses output_config.format on models with structured outputs.
• Added think, effort and fallbacks to config_Anthropic() and create_Anthropic(). With validate_model = TRUE, the config reads the model’s capabilities from the Models API.
• config_Anthropic() and create_Anthropic() default to temperature = NULL, which leaves the temperature to the model.
• An Anthropic response with stop_reason "refusal" raises an error of class rtemis_llm_refusal; a response that stops at max_tokens raises a warning.
• An OpenAI-compatible response with a refusal, including a refusal from Apple Foundation Models, raises an error of class rtemis_llm_refusal; llmapply() and map() record each refusal in the errors attribute. HTTP errors from OpenAI-compatible chat requests have class rtemis_llm_api_error and include the server’s error message.
• License changed to BSD 3-clause.
• Hashing now uses ‘openssl’, replacing the ‘digest’ dependency.
• Added image input: image_path on generate() for LLMs and Agents, and on llmapply() and agentapply(), sends local PNG, JPEG, GIF, or WebP files to Ollama, OpenAI-compatible (including Apple Foundation Models), and Anthropic-compatible backends. One prompt can be applied over many images, and several images can be sent with each prompt.
• Added chat(): a multi-turn conversation in the R console with a model name, an LLM, or an Agent, with /image, /clear, /help and /exit commands. Returns the Agent with the conversation in its memory.
• Added decision models: create_DecisionModel() with config_OllamaDecision() or config_OpenRouterDecision() fills a closed schema through generate() and dmapply(), with a probability per field from probabilities(), and answers typed questions built with choice() and noul() through decide(). as_questions() converts a closed schema into those questions, is_decidable() reports whether a schema can be filled this way, and ollama_list_decision_models() lists the decision models on an Ollama server.
• llmapply(), agentapply() and dmapply() results carry an elapsed attribute: the wall time of each call, in seconds.
• Ollama LLMs append the output schema, as JSON Schema text, to the system prompt of each request, as Ollama’s documentation recommends, so the model sees field descriptions. schema_in_prompt = FALSE on config_Ollama() and create_Ollama() sends the system prompt as written.

0.8.7

• Added config_Apple() and create_Apple() for Apple Foundation Models: Apple’s on-device model, served over the OpenAI Chat Completions wire by the rtemis-afm bridge on http://127.0.0.1:1977. AppleConfig is an OpenAIConfig, so chat, structured output, tool calling, agents, and batch generation work unchanged; no API key is resolved or sent. config_Apple() reads the bridge’s /health endpoint by default and stops with instructions when the bridge is not running or the model is unavailable (validate_model = FALSE skips the check). apple_health() returns the served model, its availability, and its context window; apple_check_available() performs the check on its own.

0.8.6

• Added items to field(): an "array" field now says what it contains, as a type name (items = "string"), a field() where the elements carry their own description or enum, or a schema() for an array of objects. Emitted as the JSON Schema items key and passed through by all three backends, so a list of values can be requested as a list instead of as one delimited string the caller splits back apart.
• An "array" field declared without items is now an error rather than a schema that fails at request time: OpenAI’s strict mode rejects an array with no element type, and a constrained-decoding backend has nothing to constrain. items on any other type is likewise refused.
• field() refuses type = "object": a field carries no properties, which OpenAI and Anthropic require of an object schema. Declare an array of objects with type = "array" and items = schema(...). tool_param() still accepts "object".
• validate_output() checks array elements against items, where before an array field constrained only its outer type.

0.8.5

• Structured responses are validated locally by default with a cached Ajv validator. validate_output = FALSE skips validation while still requesting the schema. on_validation_failure = "warn" retains invalid output and emits an informational rtemis.core message, not an R warning; "collect" is silent and "abort" raises an error carrying the response and diagnostics. Batches validate per item and summarize mismatches once at completion.
• Added standalone validate_output() and the validation_results() accessor. Reports preserve original text, input positions/names, statuses, and diagnostics, including when batch responses are extracted. Validation does not repair output.
• Agent final answers are validated before memory commit. responses() recognizes histories returned by generate(agent, ...) and extracts their final answer.
• Anthropic synthetic structured-output tool inputs are returned as answer JSON, retaining raw content metadata and excluding the synthetic tool from execution.
• Added enum to field(): restrict a field to a fixed set of permitted values. Emitted as the JSON Schema enum key and passed through unchanged by all three backends, so a backend with constrained decoding (Ollama) makes an invalid value impossible rather than merely detectable.
• Added on_error to llmapply(), agentapply(), and map(), defaulting to "na": a failed call now warns and yields NA rather than discarding every result in the batch, and the result carries an errors attribute (a data.frame of index and message) for retrying just the failures. Pass on_error = "abort" for the previous behavior of propagating the first error.
• responses() and reasoning() now accept NULL elements, mapping them to NA_character_.
• Added num_ctx and keep_alive as per-call Ollama options in generate().
• Added logprobs and top_logprobs per-call options for the Ollama and OpenAI-compatible backends, with new logprobs() and token_probs() accessors. token_probs() reads the probability of each candidate answer straight off the model’s token distribution, which is better calibrated than asking a model to emit a number. Anthropic does not return log probabilities; its messages yield NULL rather than an error.
• Per-call options passed to generate() (temperature, top_p, max_tokens, stop, top_k, seed, num_ctx, keep_alive, logprobs, top_logprobs) are validated against each backend’s documented bounds before a request is built, so an out-of-range or wrong-typed value fails locally instead of on the server. Bounds follow the backend: temperature accepts up to 2 on Ollama and OpenAI-compatible backends and up to 1 on Anthropic, and top_logprobs up to 20 on OpenAI. top_logprobs requires logprobs = TRUE, which every backend that returns alternatives needs in order to return them.

0.8.4

• Added zero_data_retention to config_OpenAI() and create_OpenAI() for per-request OpenRouter ZDR routing.
• Added think to config_Ollama(), create_Ollama(), and generate() for Ollama thinking control.

mirror server hosted at Truenetwork, Russian Federation.