render-lab-tasks-llm

Provider-agnostic LLM tasks for Render Workflows. Version 0.1.1 implements all 13 registered tasks. Version 0.1.0 contained only classification and summarization.

Install from PyPI with Python 3.12+:

pip install render-lab-tasks-llm==0.1.1

Import app, classify, classify_impl, summarize, and summarize_impl from render_lab_tasks_llm.tasks. Combine pack apps with Workflows.from_workflows. Run wrapped tasks with await ctx.run(classify, input); inject LlmDeps(chat=...) into the raw implementations for tests or custom task wrappers.

Contracts

classify accepts text, labels (objects with name and optional description), optional maxLabels (default 3), model, and ledger. It returns labels, optional reasoning, resolved model, and optional usage. Unknown/non-string labels are removed, duplicates are removed before capping, and malformed model JSON yields an empty label list, matching TypeScript. This is prompt-based classification, not a guarantee of semantic correctness.

summarize accepts text, optional instructions, maxWords (default 120), model, and ledger. It returns trimmed summary, resolved model, and optional usage. The word limit is a prompt target, not hard truncation.

Missing or null optional defaults behave like the TS implementation. Python additionally rejects negative, boolean, and nonintegral maxLabels/maxWords before calling a provider. Zero is valid. Optional result keys are omitted when unavailable. Unknown model prices do not become zero-dollar estimates.

Providers and environment

Credentials and configuration are read when a task runs, never during imports.

The default model and estimated pricing table are copied from the pinned TS snapshot, not a claim of current pricing or availability. Override LLM_MODEL for your account. The example explicitly uses openai/gpt-4o-mini.

Retries, usage, and cost tracking

Provider calls stream internally and assemble their text before returning JSON. Direct OpenAI uses max_completion_tokens; gateways/Google use max_tokens. Official SDK and HTTP retries are disabled; Render owns five retries with a 2000 ms base delay and 2x backoff. Clients close after each call.

Usage includes available input/output token counts and an estimated dollar cost when known. Output accounting preserves the TS max(completion, total-prompt) rule. Estimates ignore caching discounts and price tiers; they are not invoices.

A nonempty ledger appends a JSON cost record to Redis list llm:cost:<ledger> with TTL, compatible with the TS format. Invalid ledger configuration fails before provider spend. Pricing failures without a ledger and append/cleanup failures after spend warn without failing the task, avoiding rebilling solely to repair observability. Inject CostLedgerPort to use another store. The default Redis adapter also disables client retries.

Complete task surface

The pack also exports chat_turn (llm.chat), complete, extract, extract_strict, translate, moderate, rerank, embed, load_skill, open_cost_ledger, and cost_report, with their injectable raw implementations. with_ledger(ctx, ledger) binds the ledger while dispatching durable child runs. Chat/completion preserve stopReason so callers can detect truncated output.

LLM_EMBED_MODEL overrides the embedding default. SKILLS_PATH changes installed skill search directories. Skills load text only from GitHub, URLs, vendored files, or installed names; they never execute scripts. REDIS_URL is required only for cost-ledger IO; LLM_COST_LEDGER_TTL_SECONDS controls record expiry. Structured output uses forced tools; OpenAI-compatible calls use a non-streaming request so Gemini tool calls survive its different finish-reason behavior. Google embeddings use Google credentials, fixing the TS adapter's accidental OpenAI-key selection (tracked as a deliberate difference in ADR-0019).