---
icon: lucide/scale
---

# How piighost compares

`piighost` combines four properties a conversational agent needs. It restores the reply for the user, keeps the same placeholder over the whole conversation, hands the real value to tools, and restores while the reply streams. None of the tools below combines all four. Each one is better than `piighost` at something else, and its entry says what.

Open a solution to see how it differs from `piighost`.

??? note "Presidio (Microsoft)"

    | | `piighost` | Presidio |
    |---|---|---|
    | Detection | regex, NER or LLM | NER, regex, rules, check digits |
    | What happens to a value | reversible placeholder (memory, Redis or SQL) | mask or encrypted token |
    | Restored for the user | ✅ | ⚠️ by hand (`decrypt`) |
    | Same placeholder over the conversation | ✅ per conversation | ❌ |
    | Real value to tools, placeholder to the LLM | ✅ | ❌ |
    | Restored while streaming | ✅ | ❌ |
    | Configurable steps after detection | ✅ linking, fuzzy matching, expansion, guard rail | ⚠️ operators only |
    | Unit processed | text, conversation | text |
    | Hosting | ✅ self-hosted | ✅ self-hosted |
    | License | MIT | MIT |

    **Better at**: validating a format by its check digits, on typed text. `piighost` can also use it as a detector, with `PresidioDetector`.

??? note "LangChain PII (Python `PIIMiddleware`)"

    | | `piighost` | LangChain PII |
    |---|---|---|
    | Detection | regex, NER or LLM | regex, validators |
    | What happens to a value | reversible placeholder (memory, Redis or SQL) | mask or hash |
    | Restored for the user | ✅ | ❌ |
    | Same placeholder over the conversation | ✅ per conversation | ❌ |
    | Real value to tools, placeholder to the LLM | ✅ | ✅ |
    | Restored while streaming | ✅ | ✅ |
    | Configurable steps after detection | ✅ linking, fuzzy matching, expansion, guard rail | ❌ |
    | Unit processed | text, conversation | text, conversation |
    | Hosting | ✅ self-hosted | ✅ self-hosted |
    | License | MIT | MIT |

    **Better at**: there is nothing more to install in a LangChain agent. That is enough when the user does not need to read their real values. The JS version (`piiRedactionMiddleware`) restores the real values for the user, but not while the reply streams.

??? note "AWS Comprehend and Azure AI Language"

    | | `piighost` | AWS / Azure |
    |---|---|---|
    | Detection | regex, NER or LLM | machine learning |
    | What happens to a value | reversible placeholder (memory, Redis or SQL) | mask |
    | Restored for the user | ✅ | ❌ |
    | Same placeholder over the conversation | ✅ per conversation | ❌ |
    | Real value to tools, placeholder to the LLM | ✅ | ❌ |
    | Restored while streaming | ✅ | ❌ |
    | Configurable steps after detection | ✅ linking, fuzzy matching, expansion, guard rail | ❌ |
    | Unit processed | text, conversation | text, documents |
    | Hosting | ✅ self-hosted | ❌ cloud |
    | License | MIT | paid |

    **Better at**: models the provider maintains, to mask documents in a cloud already in place. Azure's Conversation mode only detects.

??? note "Google DLP"

    | | `piighost` | Google DLP |
    |---|---|---|
    | Detection | regex, NER or LLM | machine learning, predefined types (infoTypes) |
    | What happens to a value | reversible placeholder (memory, Redis or SQL) | stateless encrypted token |
    | Restored for the user | ✅ | ⚠️ through an API call |
    | Same placeholder over the conversation | ✅ per conversation | ✅ always the same token for a value |
    | Real value to tools, placeholder to the LLM | ✅ | ❌ |
    | Restored while streaming | ✅ | ❌ |
    | Configurable steps after detection | ✅ linking, fuzzy matching, expansion, guard rail | ⚠️ transformations |
    | Unit processed | text, conversation | text, dataset |
    | Hosting | ✅ self-hosted | ❌ cloud |
    | License | MIT | paid |

    **Better at**: transforming whole datasets in Google Cloud.

??? note "pii-redactor"

    | | `piighost` | pii-redactor |
    |---|---|---|
    | Detection | regex, NER or LLM | regex, NER |
    | What happens to a value | reversible placeholder (memory, Redis or SQL) | reversible token (vault) |
    | Restored for the user | ✅ | ✅ |
    | Same placeholder over the conversation | ✅ per conversation | ✅ per session |
    | Real value to tools, placeholder to the LLM | ✅ | ❌ |
    | Restored while streaming | ✅ | ✅ |
    | Configurable steps after detection | ✅ linking, fuzzy matching, expansion, guard rail | ❌ |
    | Unit processed | text, conversation | text, conversation |
    | Hosting | ✅ self-hosted | ✅ self-hosted |
    | License | MIT | MIT |

    **Note**: it is the closest to `piighost`. It lacks the real value to tools and the configurable steps.

??? note "Detection-only models (spaCy, GLiNER, Piiranha)"

    - Find the data without replacing or restoring it.
    - Building blocks rather than competitors. `piighost` uses them as detectors, Piiranha through `TransformersDetector`.

??? note "Dataset anonymizers (ARX, Amnesia)"

    - Transform a whole table with k-anonymity or differential privacy.
    - The result is anonymous and irreversible, where `piighost` is reversible.
    - Better at: publishing or sharing a dataset. Not made for a live conversation.

See [Limitations](limitations.md) for what `piighost` does not do, and the reasoning behind these choices.