colandix 0.1.0

ANSSI-PA-102 compliance layer for LLM inference. Runs locally. No network calls in the core path.

Colandix

Sanitize text before and after your LLM. Runs locally. No network calls in the library.

Maps to French national guidance ANSSI-PA-102 for generative AI security.

Colandix is a filtering layer that sits at the boundary between your application and the model. Safe content passes through. PII, secrets, and injection attempts do not reach the model.

---

Install

pip install colandix

Development setup with uv:

uv sync --all-groups

Optional pipelines (NER / SpaCy)

SpaCy pipelines are optional. Without them, regex, entropy, injection, and topic detectors still run; NER stays inactive.

Extra	What it installs
ner	spacy only; then python -m spacy download <pipeline> (see spacy.io/models)
ner-fr	fr_core_news_md via PyPI (matches shipped profiles)
ner-de, ner-en, ner-es, ner-it, ner-pt	spacy only; install the pipeline with python -m spacy download de_core_news_md (etc.)
ner-all	fr_core_news_md wheel only; use python -m spacy download for other languages

pip install "colandix[ner-fr]"
# or
pip install "colandix[ner]" && python -m spacy download fr_core_news_md

Bundled profiles default to fr_core_news_md. For another language, set extra.model and entities in your YAML (English uses PERSON, not PER).

---

Quickstart

from colandix import GuardPipeline

guard = GuardPipeline(profile="generique")
res_in = guard.scan_input("my prompt")

response = call_llm(res_in.sanitized_text)
guard.scan_output(response)

Tutorial: notebooks/demo_colandix.ipynb.

---

What it does

The colandix library sits between your app and the model.

flowchart LR
  userInput[User_input] --> scanIn[scan_input]
  scanIn --> sanitized[sanitized_text]
  sanitized --> llm[LLM]
  llm --> modelOutput[Model_output]
  modelOutput --> scanOut[scan_output]
  scanIn --> decisionIn["action + events"]
  scanOut --> decisionOut["action + events"]

• Scans inputs and outputs with configurable detectors (PII, secrets, entropy, prompt injection, topics, optional NER).
• Rewrites sensitive spans in sanitized_text with typed tags ([EMAIL_REDACTED], [PHONE_REDACTED], [PERSON_REDACTED], etc.).
• Decides an aggregate action: pass, warn, human_review, or block.
• Logs pseudonymized metadata; no raw prompts in logs.

---

What a scan returns

Field	Role
result.action	Aggregate decision (Action enum).
result.sanitized_text	Text to send to the LLM after masking.
result.events	Audit trail (evidence truncated for SIEM).
result.blocked	True when action == block.
guard.find_all_candidates(text)	Debug: list possible regex/NER hits.

Injection and entropy detectors affect the decision only; they do not rewrite sanitized_text.

Redaction rules: colandix/redaction.py, Architecture.

---

Profiles

Profile	Use case
generique	Everyday PII + injection.
strict	Broad PII and secrets + NER (requires SpaCy pipeline).
sante	Healthcare PII + medical topic guardrails.
dev	Credentials, secrets, sensitive dev topics.
rh	HR-sensitive wording and data.
juridique	Legal / confidentiality patterns.

YAML sources: colandix/profiles/. Trigger reference: docs/triggers-par-profil.md.

---

ANSSI PA-102 alignment

Scope: content filtering and rewriting at the model boundary.

Official guide (PDF): ANSSI generative AI security recommendations.

Ref	Topic	Coverage	Status
R25	Input/output filtering	Regex, NER, entropy, injection, sanitized_text.	Yes
R26	Application interactions	TopicDetector (allowed / blocked).	Partial
R27	Human control	Action.HUMAN_REVIEW.	Partial
R29	Logging	Structured logger, no raw user text.	Partial
R31	Critical modules	Profile separation (dev, juridique, ...).	Partial
R34	Sovereign hosting	Runs offline in the orchestration path.	Yes

Controls outside this package (TLS, IAM, org processes): docs/compliance.md.

---

LLM integration

from colandix import GuardPipeline
from colandix.exceptions import ColandixBlockedError

guard = GuardPipeline(profile="generique")

try:
    res_in = guard.scan_input(prompt, raise_on_block=True)
    response = call_llm(res_in.sanitized_text)
    guard.scan_output(response, raise_on_block=True)
except ColandixBlockedError:
    return "Content blocked."

Replace call_llm with your SDK. Use raise_on_block=False to branch on result.action yourself.

---

Custom YAML

guard = GuardPipeline(profile_path="/path/to/profile.yaml")

Start from colandix/profiles/. Detector keys: docs/architecture.md.

---

Compliance report

from colandix.compliance import print_report

print_report(guard.compliance_report())

---

Documentation

• Site (GitHub Pages)
• Architecture
• Triggers by profile
• Testing
• Compliance (ANSSI)

---

Docs (local preview)

uv run mkdocs serve     # http://127.0.0.1:8000, reloads on save
uv run mkdocs build     # static output in site/ (gitignored; CI rebuilds on deploy)

mkdocstrings requires the package to be importable at build time. uv sync --all-groups installs colandix in dev mode alongside the docs dependencies, matching what the GitHub Actions workflow does.

---

Pre-deployment checklist

uv run pytest tests/ -q
uv run ruff check colandix tests
uv run mypy colandix
uv run mkdocs build
uv build

Manual checks:

• Version in pyproject.toml matches colandix/__init__.py.
• docs/compliance.md updated if detectors or ANSSI scope changed.
• pytest tests/test_detectors/test_ner.py -v passes with fr_core_news_md installed.
• notebooks/demo_colandix.ipynb runs top-to-bottom without errors.
• docs.yml workflow completes on a push to main.

---

Publishing (maintainers)

uv build
uv publish

Set PyPI tokens and bump pyproject.toml version before publishing.

---

Contributing

1. Fork the repo.
2. git checkout -b feature/your-change
3. uv run pytest tests/ (~362 cases)
4. Open a Pull Request.

---

License

MIT — see LICENSE. © 2026 Gros Gradient

---

Pour les équipes ANSSI et security reviewers

La bibliothèque colandix est développée par Gros Gradient. La cartographie R25–R35, sanitized_text / action, et le périmètre bibliothèque sont dans docs/compliance.md (anglais). Le guide officiel ANSSI (IA générative) reste la référence réglementaire en français : PDF recommandations ANSSI.