# Model & Image Config in Sesi `model()` and `image()` take an optional config block between the model name and the prompt. Config keys are unquoted identifiers. ``` model_call := 'model' '(' string ')' config_block? '{' prompt '}' image_call := 'image' '(' string ')' config_block? '{' prompt '}' config_block := '{' key ':' value (',' key ':' value)* '}' ``` --- ## Basic Call (No Config) ```sesi let response = model("gemini-3.5-flash") {"Say hello"} show response ``` --- ## `model()` Config Keys | Key | Type | Description | | --------------- | ------------------------- | ------------------------------------------------------------ | | `thinkingLevel` | `string` | Reasoning effort: `"minimal"`, `"low"`, `"medium"`, `"high"` | | `max_tokens` | `number` | Maximum tokens in the response | | `images` | `string \| array` | Local file path(s) for vision input | | `stream` | `bool \| fn` | Stream output to stdout (`true`) or a callback fn | | `cache` | `bool` | Set to `false` to bypass Sesi Logic Caching | | `system` | `string` | System instruction for Gemini, GPT, and local models | | `search` | _(no value)_ | Enable web search grounding for real-time information | | `temperature` | `number` | ⚠️ Deprecated in Gemini 3.5+. Use `thinkingLevel`. | | `top_k` | `number` | ⚠️ Deprecated in Gemini 3.5+. Use `thinkingLevel`. | | `top_p` | `number` | ⚠️ Deprecated in Gemini 3.5+. Use `thinkingLevel`. | | `tools` | `array` | Function schemas exposed to models with tool-aware templates | --- ## Local Models Use `"local"` to run the default quantized instruction model directly: ```sesi let answer = model("local") {max_tokens: 256, temperature: 0.3} {"Explain closures simply."} ``` The first call downloads the ONNX weights to `~/.cache/sesi/models`. The loaded pipeline is reused for later calls in the same process, and downloaded weights are reused across processes. Configure the provider with environment variables: | Variable | Default | | -------------------------- | ----------------------------------------------- | | `SESI_LOCAL_MODEL` | `onnx-community/Qwen2.5-0.5B-Instruct` | | `SESI_LOCAL_DTYPE` | `q4` | | `SESI_LOCAL_DEVICE` | `cpu` | | `SESI_LOCAL_CACHE_DIR` | `~/.cache/sesi/models` | | `SESI_LOCAL_SYSTEM_PROMPT` | Sesi's concise local-assistant system prompt | | `SESI_LOCAL_WARN_TOKENS` | `2048` | An explicit model can also be selected in the model name: ```sesi let answer = model("local:onnx-community/Qwen3-0.6B-ONNX") {"Say hello."} ``` Local calls support text generation, streaming, logic caching, token usage, and tool calling for models whose chat templates support tools. Images, audio, and search grounding are rejected with an explicit error. The `system` config key overrides `SESI_LOCAL_SYSTEM_PROMPT` for one call. ```sesi let tools = [{ "type": "function", "function": { "name": "lookup_weather", "description": "Get weather by city", "parameters": { "type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"] } } }] let call = model("local:onnx-community/Qwen3-0.6B-ONNX") {tools: tools} {"Use lookup_weather for New York City."} ``` When the model requests a tool, Sesi returns canonical JSON containing `name`, `args`, and an optional `call_id`, matching hosted-model behavior. Inputs above 2,048 tokens are allowed but emit a CPU-performance warning. See [Local Models](../../docs/LOCAL_MODELS.md) for context limits, configuration, and a reference benchmark. ```sesi model("local") {max_tokens: 256, temperature: 0.5, system: "Use the tone of a motivational speaker that doesn't let clients give up on their tasks."} {query} ``` --- ## `thinkingLevel` Controls how much reasoning effort the model applies before responding: ```sesi // Fastest — minimal reasoning let r1 = model("gemini-3.5-flash") {thinkingLevel: "minimal"} {"Summarize in one sentence:" text} // Balanced let r2 = model("gemini-3.5-flash") {thinkingLevel: "low"} {"Analyze this code:" code} // Deep reasoning let r3 = model("gemini-3.5-flash") {thinkingLevel: "medium"} {"Solve this step by step:" problem} ``` --- ## `max_tokens` Cap the response length: ```sesi let brief = model("gemini-3.1-flash-lite") {max_tokens: 100} {"Explain quantum computing."} ``` --- ## `images` — Vision Input Pass a local image path to give the model visual input: ```sesi // Single image let description = model("gemini-3-flash-preview") {images: "photo.png"} {"Describe what you see."} // Multiple images let comparison = model("gemini-3.5-flash") {images: ["before.png", "after.png"]} {"What changed between these two images?"} ``` --- ## `stream` Stream tokens as they arrive instead of waiting for the full response. ### To stdout ```sesi let response = model("gemini-3.1-flash-lite") {stream: true} {"Write a short poem."} // tokens show to terminal in real-time show "Final:" response ``` ### To a callback function ```sesi fn handleChunk(chunk) { show "Chunk:" chunk } let response = model("gemini-3.1-flash-lite") {stream: handleChunk} {"Explain closures."} show "Final:" response ``` The return value is always the fully accumulated response string, regardless of whether streaming is on. --- ## `search` — Web Search Grounding `search` takes no value. Adding it to the config block tells the model to ground its response in live web search results: ```sesi let response = model("gemini-3.1-flash-lite") {search} {"What is the weather in Tokyo right now?"} show response ``` Combine with other keys normally: ```sesi let response = model("gemini-3.1-flash-lite") {search, max_tokens: 200} {"Latest news in Media this week."} ``` --- ## `cache` Sesi caches model responses by default. Set `cache: false` to force a fresh call: ```sesi let fresh = model("gemini-3-flash-preview") {cache: false} {"What time is it?"} ``` --- ## Combining Config Keys Multiple keys are comma-separated on one line: ```sesi let result = model("gemini-3.5-flash") {thinkingLevel: "medium", max_tokens: 500} {"Analyze this document:" doc} let scan = model("gemini-3.5-flash") {images: "receipt.png", thinkingLevel: "minimal"} {"Extract all line items as JSON."} ``` --- ## `image()` Config Keys | Key | Type | Description | | -------- | ------------------------- | --------------------------------------------------- | | `ratio` | `string` | Aspect ratio — e.g. `"1:1"`, `"16:9"`, `"4:3"` | | `size` | `string` | Output resolution — `"512"`, `"1K"`, `"2K"`, `"4K"` | | `images` | `string \| array` | Reference image(s) for style/context | ```sesi let logo = image("gemini-3.1-flash-image") {ratio: "1:1", size: "512"} {"A minimal logo for a programming language"} write_image("logo.png", logo) ``` ```sesi let banner = image("gemini-2.5-flash-image") {ratio: "16:9", size: "1K"} {"A dark futuristic cityscape at night"} write_image("banner.png", banner) ``` --- ## Quick Reference ```sesi // No config let r = model("gemini-3.5-flash") {"Hello"} // thinkingLevel let r = model("gemini-3.5-flash") {thinkingLevel: "low"} {"Summarize:" text} // max_tokens let r = model("gemini-3.5-flash") {max_tokens: 200} {"Explain this."} // Vision input let r = model("gemini-3.5-flash") {images: "scan.png"} {"Transcribe all text."} // Streaming to stdout let r = model("gemini-3.1-flash-lite") {stream: true} {"Write a poem."} // Streaming with callback fn onChunk(chunk) { show chunk } let r = model("gemini-3.1-flash-lite") {stream: onChunk} {"Tell a story."} // No cache let r = model("gemini-3.1-flash-lite") {cache: false} {"What's trending?"} // Combined let r = model("gemini-3.5-flash") {thinkingLevel: "medium", max_tokens: 500, images: "doc.png"} {"Analyze this."} // image() let img = image("gemini-2.5-flash-image") {ratio: "16:9", size: "1K"} {"A sunset over the ocean"} write_image("output.png", img) ``` ---