iil-aifw 0.11.5

Django AI Services Framework — DB-driven LLM routing, quality tiers, NL2SQL

iil-aifw — Django AI Services Framework

DB-driven LLM provider, model & usage management for Django projects.

Installation

pip install iil-aifw
# With NL2SQL support:
pip install "iil-aifw[nl2sql]"

Extras / Optional Dependencies

Extra	Additional deps	Purpose
nl2sql	(none — activation-only)	Activates aifw.nl2sql INSTALLED_APPS entry + migrations; no additional download weight
rag	pgvector>=0.3	Schema-RAG via pgvector (ADR-009 Phase 1, planned)
dev	pytest, hatch, ruff, iil-promptfw	Development toolchain

Quick Start

# settings.py
INSTALLED_APPS = [
    ...
    "aifw",
]

# Run migrations
python manage.py migrate aifw
python manage.py init_aifw_config   # seed default providers & models

from aifw.service import sync_completion

result = sync_completion(
    action_code="story_writing",
    messages=[{"role": "user", "content": "Write a short story about a dragon."}],
)

if result.success:
    print(result.content)
else:
    print(result.error)

Features

• DB-driven model routing — swap LLM providers/models via Django Admin, zero code changes
• Multi-provider — OpenAI, Anthropic, Google, Ollama, any LiteLLM-compatible provider
• Quality-level routing — map subscription tiers to model quality (economy/balanced/premium)
• Async & sync — completion() async, sync_completion() sync, completion_with_fallback()
• Streaming — sync_completion_stream() for Django StreamingHttpResponse
• Usage logging — automatic token & latency tracking per action, per tenant
• Fallback models — configure primary + fallback model per action type
• NL2SQL — natural language → SQL engine with few-shot examples and self-healing retry
• Hybrid cache — process-local (30s) + Django cache framework (600s, Redis-aware)

Core API

from aifw.service import (
    completion,                  # async
    sync_completion,             # sync wrapper
    completion_with_fallback,    # async, tries fallback model on error
    sync_completion_with_fallback,
    sync_completion_stream,      # streaming generator
    get_quality_level_for_tier,  # tier name → QualityLevel int
    check_action_code,           # bool — action code exists in DB
)

Quality-Level Routing (v0.10.2)

Each AIActionType row can be scoped to a quality_level (1–9) and/or priority ('fast'|'balanced'|'quality'). A NULL in either dimension acts as catch-all. Lookup uses a deterministic 4-step cascade: exact → ql-only → prio-only → catch-all.

from aifw.constants import QualityLevel
from aifw.service import get_quality_level_for_tier, sync_completion

# Map a subscription tier to a quality level (DB-driven via TierQualityMapping)
ql = get_quality_level_for_tier("pro")          # → QualityLevel.BALANCED (5)
ql = get_quality_level_for_tier("premium")      # → QualityLevel.PREMIUM  (8)
ql = get_quality_level_for_tier("freemium")     # → QualityLevel.ECONOMY  (2)

result = sync_completion(
    "story_writing",
    messages,
    quality_level=ql,
    priority="quality",
)

Multi-Tenant Usage Tracking (v0.5.0)

result = sync_completion(
    "story_writing",
    messages,
    tenant_id=user.organization_id,
    object_id=f"story-{story.pk}",
    metadata={"source": "web", "plan": "pro"},
)

Privacy-by-Design Logging (issue #8)

AIUsageLog can pseudonymise or anonymise itself at write time — PII never reaches the database. Select the policy with one Django setting (DSGVO Art. 25 / Art. 5 Abs. 1 lit. c):

# settings.py
from aifw.constants import PrivacyMode
AIFW_PRIVACY_MODE = PrivacyMode.PSEUDONYMOUS   # "full" (default) | "pseudonymous" | "anonymous"
AIFW_PRIVACY_HMAC_SECRET = env("AIFW_PRIVACY_HMAC_SECRET")  # optional; falls back to SECRET_KEY

# Call site stays the same — the hook transforms the row before it is written:
sync_completion(
    "nl2sql", messages,
    user=request.user,
    tenant_id=user.organization_id,
    metadata={"nl_question": question},
)

Mode	user	metadata
full (default)	stored raw (FK)	stored raw
pseudonymous	NULL	user_hash = HMAC(user.pk); nl_question → topic
anonymous	NULL	reduced to {"day_bucket": "<ISO date>"} (tenant + tokens survive)

The privacy_mode column records which policy was applied to each row.

Topic classification is a plain callable — aifw never imports iil-promptfw. The built-in pseudonymous classifier emits a coarse placeholder ("unclassified"); wire a real classifier via a custom hook:

# myapp/privacy.py
from aifw.privacy import PseudonymousHook

def my_hook():
    from promptfw import classify_topic
    return PseudonymousHook(topic_classifier=classify_topic)

# settings.py
AIFW_PRIVACY_HOOK = "myapp.privacy:my_hook"   # dotted path → PrivacyHook / subclass / factory

A custom hook receives the create-payload dict (keys: user, tenant_id, metadata, …) and returns it rewritten. If a configured non-full hook raises, logging fails closed — user/metadata are scrubbed rather than leaking PII.

For supervisor-facing reports, AIUsageLog.objects.aggregate_with_k_anonymity( *group_by, k=3) returns only buckets with ≥ k entries (suppresses re-identifiable small groups).

Default-shift roadmap: the default stays "full" in the 0.11.x line to keep existing audit views (bfagent, weltenhub, travel-beat) working unchanged. A shift to "pseudonymous" as the default is planned for 0.12.0 as a documented breaking change.

Streaming (v0.4.0)

from aifw.service import sync_completion_stream
from django.http import StreamingHttpResponse

def stream_story(request):
    def generate():
        for chunk in sync_completion_stream("story_writing", messages):
            yield chunk
    return StreamingHttpResponse(generate(), content_type="text/plain")

Django Models

Model	Purpose
LLMProvider	Provider config (API key env var, base URL)
LLMModel	Model config (max tokens, cost per million tokens)
AIActionType	Action → model mapping; supports quality_level + priority dimensions
AIUsageLog	Token/latency/cost tracking per request, with tenant_id + metadata + privacy_mode
TierQualityMapping	Subscription tier → quality_level mapping (DB-driven)

NL2SQL (v0.7.0+)

# settings.py
INSTALLED_APPS = [
    ...
    "aifw",
    "aifw.nl2sql",   # activates NL2SQL models + migrations
]

python manage.py migrate aifw_nl2sql
python manage.py seed_nl2sql_examples   # seed verified few-shot examples

from aifw.nl2sql import NL2SQLEngine

engine = NL2SQLEngine(schema_xml="<schema>...</schema>")
result = engine.query("How many open orders do we have?")

if result.success:
    print(result.sql)
    print(result.rows)

• Few-shot examples injected automatically from NL2SQLExample model
• Self-healing retry — on SQL execution error, a second LLM call includes the error as context
• Feedback capture — NL2SQLFeedback auto-created on every execution error
• Management commands: seed_nl2sql_examples, promote_feedback, validate_schema

Management Commands

python manage.py init_aifw_config       # seed default providers & models
python manage.py check_aifw_config      # CI gate — verify all action codes have a catch-all row
python manage.py seed_nl2sql_examples   # seed NL2SQL few-shot examples
python manage.py promote_feedback       # promote corrected feedback → NL2SQLExample
python manage.py validate_schema        # validate schema-XML against real DB (exit 1 on error)

Cache Tuning

# Environment variables (optional)
AIFW_LOCAL_CACHE_TTL=30     # process-local TTL in seconds (default: 30)
AIFW_CACHE_TTL=600          # shared cache TTL in seconds  (default: 600)

from aifw.service import invalidate_action_cache, invalidate_tier_cache

invalidate_action_cache("story_writing")   # clear one action's cache entries
invalidate_action_cache()                  # clear all
invalidate_tier_cache("pro")               # clear one tier

Constants

from aifw.constants import QualityLevel

QualityLevel.ECONOMY   # 2
QualityLevel.BALANCED  # 5
QualityLevel.PREMIUM   # 8
QualityLevel.ALL       # (2, 5, 8)

Configuration

Django Settings

INSTALLED_APPS = [
    "aifw",           # core: LLMProvider, LLMModel, AIActionType, AIUsageLog, TierQualityMapping
    "aifw.nl2sql",    # optional: NL2SQL models + migrations (requires `nl2sql` extra)
]

Run once after installation to seed default providers and models:

python manage.py migrate aifw
python manage.py init_aifw_config

Environment Variables (optional)

Variable	Default	Purpose
AIFW_LOCAL_CACHE_TTL	30	Process-local cache TTL in seconds
AIFW_CACHE_TTL	600	Shared cache TTL in seconds (Redis-backed)
AIFW_PRIVACY_MODE	full	Privacy policy for AIUsageLog: full | pseudonymous | anonymous (issue #8)
AIFW_PRIVACY_HOOK	—	Dotted path to a custom PrivacyHook (overrides AIFW_PRIVACY_MODE)
AIFW_PRIVACY_HMAC_SECRET	SECRET_KEY	HMAC key for pseudonymous user_hash

AIFW_PRIVACY_* are read from Django settings (not env), like Django's own SECRET_KEY. The cache TTLs above are read from the process environment.

No Django settings dict is required — all configuration is managed via Django Admin models.

Changelog

See CHANGELOG.md.