lexigram-ai-feedback

AI feedback collection for the Lexigram Framework — collection, processing, and storage

Overview

AI feedback collection and continuous-learning loop for the Lexigram Framework. Captures user ratings, corrections, text feedback, and ground-truth labels from LLM interactions and routes them through an extensible processor pipeline to configurable storage backends. Zero-config usage starts with sensible defaults.

Full documentation: docs.lexigram.dev

Install

uv add lexigram-ai-feedback

Quick Start

from lexigram import Application
from lexigram.di.module import Module, module

from lexigram.ai.feedback import FeedbackModule
from lexigram.ai.feedback.config import FeedbackConfig


@module(
    imports=[
        FeedbackModule.configure(
            FeedbackConfig(
                enabled=True,
                async_processing=True,
                store_raw_payloads=False,
            )
        )
    ]
)
class AppModule(Module):
    pass


async with Application.boot(modules=[AppModule]) as app:
    # use app.container to resolve services
    ...

Configuration

Zero-config usage: Call FeedbackModule.configure() with no arguments to use defaults.

Option 1 — YAML file

# application.yaml
ai_feedback:
  enabled: true
  async_processing: true
  store_raw_payloads: false

Option 2 — Profiles + Environment Variables (recommended)

export LEX_AI_FEEDBACK__ENABLED=true
# Environment variables for each field

Option 3 — Python

from lexigram.ai.feedback.config import FeedbackConfig
from lexigram.ai.feedback import FeedbackModule

config = FeedbackConfig(
    enabled=True,
    async_processing=True,
    store_raw_payloads=False,
)
FeedbackModule.configure(config)

Config reference

Module Factory Methods

Key Features

• Four feedback types: Rating, free-text, correction (original → corrected), and ground-truth labels
• Extensible processor pipeline: Custom processors via FeedbackProcessorRegistry
• Storage backends: In-memory, database (DatabaseFeedbackStore), and cache (CachedFeedbackStore)
• Middleware integration: FeedbackMiddleware and FeedbackContext for request/response capture
• Lifecycle hooks: FeedbackSubmittedHook, FeedbackProcessedHook, FeedbackStoredHook

Security

Enforced payload size limits

Submitted feedback is validated at the submission chokepoints (FeedbackCollector._store() for the middleware/processor pipeline and the collect_* API; FeedbackService.submit_feedback() for the programmatic service). Oversized payloads are rejected with FeedbackTooLargeError (never silently truncated); boundary values (exactly at the limit) pass.

Optional endpoint authorization

create_feedback_endpoint() performs no identity check by default — an endpoint mounted without an authorize callback accepts feedback from anyone who can reach it; that is an explicit, informed choice, not an enforced control. Pass an authorization callback to gate submissions:

from lexigram.ai.feedback import FeedbackCollector, FeedbackMiddleware


def authorize(ctx) -> bool:
    # ctx.context_id is the context being submitted against;
    # ctx.metadata carries what the host framework supplied (e.g. user)
    return ctx.metadata.get("user_id") is not None


middleware = FeedbackMiddleware(
    collector=FeedbackCollector(),
    authorize=authorize,
)
app.post("/feedback", middleware.create_feedback_endpoint())

Sync and async (bool | Awaitable[bool]) callables are supported. A denied submission raises FeedbackAuthorizationError before any processing.

Testing

async with Application.boot(modules=[FeedbackModule.stub()]) as app:
    # your test code
    ...

Key Source Files