What silicon panels are for | Panels and instruments | Compare responses with a benchmark | Conjoint designs | A live administration | Request counts and model choice | Relations

A first agent | Tools with declared powers | Budgets: call ceilings and recorded-use gates | Structured answers | Agents calling agents | Pipelines: a fixed sequence of specialists | A two-agent conversation

State in, state out | Branches and loops | Checkpoints, resume, and fork | Replay verification | When to reach for this

Tools: the model consults your R session | Streaming: watch the reply arrive | Logprobs: the model's confidence as data | Related features

Prepare the Text Data | Configure Embedding Model | Generate Embeddings | Analyze Similarity | Expected Results

DeepSeek: deepseek-chat | DeepSeek: deepseek-reasoner | Groq: llama-3.1-8b-instant | Groq: openai/gpt-oss-20b | Using the chat history | Inspect the full conversation | Structured chat in one call (DeepSeek example)

Overview | Build a factorial design | Run unstructured | Structured version

llm_fn: unstructured (DeepSeek) | llm_fn: unstructured (Groq) | llm_fn_structured: schema-first (DeepSeek) | llm_mutate: unstructured (Groq) | llm_mutate: shorthand syntax | llm_mutate with .structured flag | Soft structured output with tags | Row batching: many rows per call | Preview before you spend, summarize after | llm_mutate_structured: structured with shorthand (Groq)

Analyze simulated and imported transcripts | Run a focus group | Run descriptive analysis offline | Opt in to model analysis | Import and analyze an existing transcript | Construct the objects directly | Interpretation

Architecture | Main functions | Offline execution and replay | Running a session | Messages and prompts | Stored data | Analysis | Extending the package

One object, five views | Diagnostics for a run | Reporting within scope | The manifest: identity without the transcript | Archiving: seal the evidence | Redaction without lying | Why archive

Robustness under perturbation | Personas are prompts with provenance | Claim types state the design's scope | A defensible sequence

Designing the Experiment | Vectorized Execution with Soft Structuring | Conclusion

Overview | What the major providers actually support | Why prefer schema output? | Why JSON-only still matters | Quirks you will hit in practice | LLMR helpers for common parsing failures | Minimal patterns (guarded code) | 1) JSON mode, no schema (works across OpenAI-compatible providers) | 2) Schema-constrained output on Groq (open model) | 3) DeepSeek: JSON-object mode with local validation | 4) Groq: another structured output example | Defensive patterns (no API calls) | Parallel execution with schema validation | Choosing the mode | References

1. An audit trail for every call | 2. Replication and reliability | 3. A methods paragraph you can edit | 4. What did it cost? | 5. Prompt caching | 6. Batch jobs at half price | 7. Seeds, model versions, and reproducibility limits

Debate: the strongest case for each side | Focus group: opinions in company | Interview: depth on one respondent | Deliberation: talk, then decide | From transcripts to analysis

A governed tool declares its powers | Guardrails check what flows in and out | Human approval for the irreversible | Putting it together

1. Install | 2. Set an API key | 3. Your first call | 4. Apply a model to a data frame | 5. Generative calls over a vector | 6. Tagged fields and row batching | 7. Embeddings | 8. Look before you spend, summarize after | Where to go next