Document feedback-aware validation retries

This commit is contained in:
2026-08-27 01:35:23 +00:00
parent 7f01c3e79e
commit b2e83bd6e7
12 changed files with 229 additions and 38 deletions

View File

@@ -132,7 +132,7 @@ model: example-model
Keep credentials out of the local-backend object. A PromptKit profile may name
its credential environment variable through `api_key_env`; set that variable
only in the run environment. PromptKit owns the
[pinned profile-file format](https://gitea.maximumdirect.net/eric/promptkit/src/tag/v0.8.0/docs/formats.md),
[pinned profile-file format](https://gitea.maximumdirect.net/eric/promptkit/src/tag/v0.9.0/docs/formats.md),
including `base_profile` inheritance. Notarius passes profiles through without
merging them. Filesystem profiles cannot express PromptKit's in-memory
`APIKeyRequired` setting; an unset `api_key_env` is optional and may reach the
@@ -252,12 +252,11 @@ explicit `null` and non-integer values are invalid. An explicit value on a
deterministic binding or validator is invalid, while a pipeline value simply
does not apply to deterministic selections.
`validation_policy` records terminal-policy metadata for a complete producer
attempt and validation chain. It may appear on a pipeline or a **chunk**,
`validation_policy` controls terminal disposition for one complete producer
attempt and validator chain. It may appear on a pipeline or a **chunk**,
**extract**, **merge**, or **normalize** module binding; input, output, and
validator bindings reject it. Every field is optional and resolves in binding,
pipeline, then application-default order. The current runner records the
resolved values but does not yet consume them for terminal disposition:
pipeline, then application-default order:
| Field | Values | Default |
| --- | --- | --- |
@@ -270,6 +269,16 @@ rejected. A deterministic producer may not explicitly set
**producer_structural_failure** on its binding, although a pipeline-level
default remains valid for pipelines that include LLM-backed producers.
After the producer binding's retry budget is exhausted, an invalid structured
response uses **producer_structural_failure**. One or more semantic validator
rejections use **semantic_rejection**; rejection takes precedence over an
exhausted validator failure or skip. With no rejection, an exhausted validator
failure or skip uses **validator_failure**. `reject_output` records the
terminal rejection without advancing that candidate. `warn_continue` is valid
only for validator execution failure: it advances a structurally valid,
otherwise unrejected result with incomplete-validation provenance and without
making it reusable checkpoint state.
A lane has these fields:
| Field | Type | Default | Rules |
@@ -309,7 +318,7 @@ extract:
| **llm_profile** | string | none | Optional non-empty PromptKit profile ID for an LLM-backed binding. It overrides the pipeline default unless the run supplies **--llm-profile**. |
| **structured_output_repair_attempts** | integer | pipeline or prompt-owned (1 in maintained production prompts) | Optional structural-repair limit from 0 through 3 for an LLM-backed binding. It overrides the pipeline value; explicit 0 disables structural repair. |
| **validation_policy** | object | pipeline or application defaults | Optional field-by-field terminal-policy override for a chunk, extract, merge, or normalize binding. |
| **retries** | integer | 0 | Non-negative additional attempts for chunk, extract, merge, and normalize bindings. |
| **retries** | integer | 0 | Non-negative additional complete producer attempts for chunk, extract, merge, and normalize bindings. This single budget covers operational errors, invalid structured output, module-requested normalization retry, and semantic correction. |
| **options** | object | none | Must satisfy the selected module. |
| **references** | map | none | Valid only on chunk, extract, merge, and normalize bindings. |
| **validators** | list | production chain | Valid only on chunk, extract, merge, and normalize bindings. |
@@ -318,13 +327,22 @@ Omitting **validators** uses the registered chain. **validators: []** selects
an empty chain; a non-empty list replaces the chain in the listed order.
Validator bindings accept only **module**, **llm_profile**,
**structured_output_repair_attempts**, **retries**, and **options**. Their
**retries** value is a non-negative validator-execution budget and is valid
only when the selected validator is LLM-backed. They reject
**retries** value is a non-negative additional validator-execution budget and
is valid only when the selected validator is LLM-backed. A validator retry
rechecks the same immutable candidate; it never regenerates the producer.
They reject
**validation_policy**, **references**, and nested **validators**. Deterministic
validators reject explicit **llm_profile** and
**structured_output_repair_attempts**.
Deterministic module bindings also reject those explicit fields.
An LLM-backed chunk, extract, merge, or normalize producer with both a
non-empty validator chain and positive **retries** must declare the supported
single-response correction capability. Preparation rejects a configuration
that could require semantic correction from a producer that cannot provide an
exact prior response. A deterministic producer, or an LLM attempt that did
not make a model call, cannot consume a semantic retry after rejection.
The **json** output module accepts optional **include_chunk_map** and
**evidence_context** settings: