Configuration Reference

On this page

PromptEngine(...)
PromptEngine.build(...) / PromptRequest
ContextManager(...)
GuardrailEngine(...)
PromptOptimizer(...)
ABTestRunner(...)
Enum-valued options at a glance
Best practices

Every configurable option in the library, in one place. All of these are plain Python constructor/keyword arguments — there is no config-file loader, environment-variable reader, or CLI flag parser anywhere in the source.

`PromptEngine(...)`

Option	Type	Default	Purpose
`context_manager`	`ContextManager`	new instance	Override context-engineering behavior
`guardrail_engine`	`GuardrailEngine`	new instance	Override guardrail selection/rendering
`extra_guardrails`	`List[Guardrail]`	`None`	Add custom always-considered guardrails
`enable_cache`	`bool`	`True`	Toggle the SHA-256 prompt cache
`cache_ttl_sec`	`int`	`300`	Cache entry lifetime
`max_cache_size`	`int`	`256`	Max cache entries before oldest-eviction
`enable_auto_detect`	`bool`	`True`	Toggle keyword-based type/strategy/complexity inference
`memory_store`, `cache_store`, `router`	`Any`	`None`	Stored but unused by any method in the provided source — see architecture

See api/prompt_engine.

`PromptEngine.build(...)` / `PromptRequest`

Option	Type	Default	Runtime effect
`prompt_type`	`PromptType \| str`	`QA`	Selects the default strategy (via `PromptBuilder`) and gates `STAY_ON_TOPIC` guardrail
`strategy`	`PromptStrategy \| str`	`SIMPLE`	Selects the concrete `BaseStrategy` used
`output_format`	`OutputFormat \| str`	`TEXT`	Selects the format instruction appended by `GuardrailEngine`
`complexity`	`QueryComplexity \| str`	`SIMPLE`	May trigger a strategy upgrade (`SIMPLE` type default → `COT`/`MULTI_HOP`) and adds `SELF_CHECK` guardrail at `COMPLEX`/`EXPERT`
`user_profile`	`UserProfile \| str`	`GENERAL`	Selects the persona line appended to the system prompt
`max_context_tokens`	`int`	`3000`	Token budget passed to `ContextManager.build`
`max_answer_tokens`	`int`	`512`	Stored on `PromptRequest`; not read by any code path in the provided source — appears to be a hint intended for the calling application (e.g. to set the LLM API's own `max_tokens`), not enforced internally
`output_schema`	`Optional[Dict]`	`None`	Appended as a JSON schema block when `output_format == STRUCTURED`
`language`	`str`	`"en"`	Resolves to `Lang.EN`/`Lang.AR` templates via `resolve_lang`
`enable_guardrails`	`bool`	`True`	Gates most (not all) guardrails — see api/guardrails.md
`enable_citations`	`bool`	`True`	Gates the `CITE_SOURCES` guardrail (also requires non-empty `documents`)
`enable_uncertainty`	`bool`	`True`	Gates the `UNCERTAINTY` guardrail
`session_id`, `user_id`, `trace_id`	`str`	`""`	Passed through to `BuiltPrompt` and the trace log; not validated
`extra`	`Dict`	`{}`	Only key read anywhere: `extra["tools"]`, by `ReActStrategy`

[!WARNING] max_answer_tokens is not enforced Despite being documented in PromptEngine.build's own docstring as "a hint to the LLM about expected answer length", no code path in the provided source reads request.max_answer_tokens after it's stored. It is not injected into any prompt text and not used to truncate anything. If you need the model to respect an answer-length limit, you must communicate it yourself (e.g. via the LLM API's own max_tokens parameter, or by adding a custom guardrail).

`ContextManager(...)`

Option	Type	Default	Runtime effect
`dedup_threshold`	`float`	`0.85`	Jaccard similarity cutoff for near-duplicate removal
`min_doc_tokens`	`int`	`5`	Minimum approximate tokens for a doc to be considered
`use_lost_in_middle`	`bool`	`True`	Enables relevance-aware reordering (docs > 3)
`summarize_overflow`	`bool`	`False`	No effect — declared but never read (see api/context_manager)
`max_memory_messages`	`int`	`10`	Default conversation turns (not messages) kept by `format_memory`

`GuardrailEngine(...)`

Option	Type	Default	Runtime effect
`extra_guardrails`	`List[Guardrail]`	`None`	Always appended after rule-based selection, regardless of `enable_guardrails`

`PromptOptimizer(...)`

Option	Type	Default	Runtime effect
`max_total_tokens`	`int`	`6000`	Hard cap; only `user` is ever truncated to meet it
`enable_filler`	`bool`	`True`	Strip common LLM padding phrases
`enable_dedup`	`bool`	`True`	Remove `user` paragraphs duplicated verbatim in `system`
`enable_whitespace`	`bool`	`True`	Normalize spaces/blank lines

`ABTestRunner(...)`

Option	Type	Default	Runtime effect
`llm`	`Any`	required	Must expose `.generate(str) -> str`
`variants`	`List[PromptVariant]`	required, ≥ 2	Raises `ValueError` if fewer than 2
`rag_system`	`Any`	`None`	Optional automatic context retrieval
`judge_llm`	`Any`	`None` (→ `llm`)	Separate model for scoring
`batch_delay_sec`	`float`	`0.0`	Sleep between queries in `run()`

Enum-valued options at a glance

See api/types for the full member tables of PromptType, PromptStrategy, OutputFormat, QueryComplexity, and UserProfile.

Best practices

Set enable_cache=False in any environment where prompt inputs vary in ways not captured by the cache key (see api/prompt_engine.md#cache) — e.g. if document metadata or scores matter to your use case but aren't part of the hash.
Don't rely on max_answer_tokens to constrain model output length; it is currently inert.
If you need summarize_overflow-style behavior (summarizing docs that don't fit the budget instead of truncating them), you'll need to implement it yourself — the flag exists but does nothing in this version.

Configuration Reference

PromptEngine(...)

PromptEngine.build(...) / PromptRequest

ContextManager(...)

GuardrailEngine(...)

PromptOptimizer(...)

ABTestRunner(...)