Basic Imp program that maps signature inputs to typed outputs with an LM.
Predict is the smallest executable Imp module. It formats a signature and
input map with an adapter, calls the configured LM, parses the result into a
Imp.Prediction, and attaches trace metadata. Most higher-level modules
such as ChainOfThought, RAG, ReAct, BestOfN, and optimizers eventually compose
around this shape, which is why they live under Imp.Predict.*: as in DSPy,
where dspy.ChainOfThought wraps a dspy.Predict, each of them holds or
builds one.
Required inputs are validated before an LM call is made. This keeps missing data as a local program error instead of spending provider calls on malformed prompts.
Example
iex> lm = Imp.LM.Static.new(handler: fn _messages, _opts -> %{answer: "4"} end)
iex> program = Imp.Predict.new("question -> answer", lm: lm)
iex> {:ok, prediction} = Imp.Predict.call(program, %{question: "2+2?"})
iex> Imp.Prediction.get(prediction, :answer)
"4"
iex> silent_lm = Imp.LM.Static.new(handler: fn _messages, _opts -> %{answer: "unused"} end)
iex> missing = Imp.Predict.new("question, context -> answer", lm: silent_lm)
iex> Imp.Predict.call(missing, %{question: "2+2?"})
{:error, {:missing_input_fields, [:context]}}Summary
Functions
Calls the program with a map or keyword list of inputs.
Calls the program with a per-call config override.
Serializes portable program state for Imp.Saving.
Builds a prediction program.
Returns a copy of the program with demonstrations attached.
Returns a copy of the program pinned to a concrete LM.
Returns a copy of the program with a new signature.
Functions
Calls the program with a map or keyword list of inputs.
Returns {:ok, prediction} on success or {:error, reason} for local input
errors, LM errors, or adapter parse errors. Successful predictions include
redacted trace metadata with the rendered messages and raw LM output.
A completion that could not be read as the outputs, after the chat and XML
adapters' JSON fallback, is {:error, %Imp.AdapterParseError{}} with its
:trace set; see that module for the kinds. An LM request that failed,
including the fallback's own request, is the LM's error (Imp.LMError for
Imp.Clients.ReqLLM).
Calls the program with a per-call config override.
config is a keyword list merged over the program's stored config for this
invocation only; the program itself is not mutated. Every merged entry flows
to the LM request, including a predicted-outputs :prediction map. Raises
ArgumentError if config is not a keyword list.
Imp.Predict.call(program, %{question: "..."},
temperature: 0.2,
prediction: %{type: "content", content: "..."}
)Serializes portable program state for Imp.Saving.
Builds a prediction program.
Pass :lm and :adapter for a self-contained program, or omit them to resolve
from process/global Imp settings at call time. An unknown option raises
ArgumentError.
Options
:lm- The model the program calls. When absent, each call usesImp.Settings':lm.:adapter- The adapter that renders the request and parses the reply. When absent, each call usesImp.Settings':adapter.:demos(list ofterm/0) - Worked examples rendered before the inputs, asImp.Examplevalues or maps. The default value is[].:config(keyword/0) - Request options sent with every call: sampling (:temperature,:max_tokens,:n, ...) and the other options the LM client takes. The default value is[].:adapter_opts(keyword/0) - Options handed to the adapter'sformat/3on every call, beside:demos: how a program passes rendering data, and how a host injects renderers without writing a second adapter module. The default value is[].:metadata(map ofterm/0keys andterm/0values) - Free-form metadata kept on the program. The default value is%{}.
Returns a copy of the program with demonstrations attached.
Returns a copy of the program pinned to a concrete LM.
Returns a copy of the program with a new signature.