# Philter UCSF Guide

A source-qualified guide to when Philter UCSF fits the locally deployable text redaction job.

Philter UCSF is eligible for the clinical-note subset of local text de-identification; a specialized alternative, not a general-language drop-in.

Site: Field Guides
Category: Software
Version: 2026-10-18.1
Published: 2026-10-18
Updated: 2026-10-18
Freshness: Checked 2026-10-18 against the cited sources. Official product claims, independent review/community signals, and evidence limits are reported; no hands-on benchmark, installation test, or privacy audit is claimed.
Tags: software, privacy, pii, redaction, philter-ucsf
Canonical: https://fieldguides.co/g/philter-ucsf
Markdown: https://fieldguides.co/g/philter-ucsf.md
Versioned Markdown: https://fieldguides.co/g/philter-ucsf/2026-10-18.1.md
JSON: https://fieldguides.co/api/guides/philter-ucsf

## Quick Read

- Verdict: Choose it for clinical-note filtering. - Philter UCSF fits the clinical-note subset of local text de-identification. It is a specialized alternative rather than a general-language drop-in.
- Workflow: Remove protected health information from clinical text. - Configure the filters and input/output paths for clinical plain-text notes, then inspect the redacted output. Evaluate it on representative clinical text before extending the workflow to another domain.
- Boundary: Check the limit before committing - The key constraint is Clinical-domain focus; general chat/email quality is not established. No source here proves a universal performance, privacy, or reliability outcome.

## Decision

**Choose Philter UCSF when its documented boundary matches the job.**

Philter UCSF fits the clinical-note subset of local text de-identification. The cited sources describe it as a locally deployable redaction option. Choose it when clinical-note de-identification is the job. Its clinical-domain focus does not establish quality for general chat or email.

## Setup

**Start by confirming deployment and the data path.**

Confirm this setup detail before you build a routine: Local Python command-line package. Then verify the processing boundary — Configurable filtering pipeline with optional evaluation annotations — and keep the first sample low-risk until the account, device, or runtime behavior matches the source description.

## Workflow

**Workflow: remove protected health information from clinical plain text.**

Use Philter UCSF to remove protected health information from clinical plain text. Customization is described as filter configuration and input/output paths. Keep the workflow small enough to review its output and control settings; the cited sources do not establish a universal quality or speed result.

## Alternatives

**Compare Philter UCSF by the constraint that decides the work.**

The five-guide table retains the current guide and four siblings in the locally deployable text redaction group. Values come from the cited source cells; an Unknown cell means the sources did not support a stronger statement.

| Guide | Workflow | Deployment | Input | Entry cost | Method | Customization | Limitation |
| --- | --- | --- | --- | --- | --- | --- | --- |
| OpenAI Privacy Filter | Model-based text redaction with a local CLI and Python API. | Local CPU or GPU; checkpoints downloaded if absent. | Text strings, files and pipes. | Open repository; compute and operational costs remain. | Token classification with span decoding. | Runtime precision/recall presets, evaluation and fine-tuning commands. | Validate detection on the target corpus; community scores change with matching criteria. |
| Presidio | PII identification followed by configurable anonymization. | Python module or self-hosted HTTP service/container. | Text and detected entity spans. | Open-source deployment; self-hosting/integration cost separate. | Analyzer plus separate anonymizer operators. | Replace, mask, redact or encrypt; reversible operators need corresponding metadata and keys. | Anonymizer acts on supplied detections; missing detections are not repaired by replacement. |
| scrubadub | Replace personally identifiable spans in free text. | Python package invoked inside local scripts. | Free-text strings. | Apache-2.0 software; dependency and operation costs separate. | Detectors identify spans; postprocessors control replacement. | Optional address, spaCy and Stanford detector packages; locale configuration. | Extra detector dependencies are needed for some entity types. |
| Limina on-premises | Detect and de-identify sensitive data; this comparison covers text. | Container in customer on-premises environment or VPC. | Text; product also supports documents, images and audio. | Enterprise custom pricing; free hosted Starter is not the local-deployment price. | Commercial entity detection and de-identification engine. | Enterprise model-version pinning and guided integration/tuning. | Local availability and commercial terms require vendor agreement. |
| Philter UCSF | Remove protected health information from clinical plain text. | Local Python command-line package. | Plain-text file or directory of clinical notes. | BSD-3-Clause software; integration and compute separate. | Configurable filtering pipeline with optional evaluation annotations. | Filter configuration and input/output paths. | Clinical-domain focus; general chat/email quality is not established. |

## Review signal

**Independent context for Philter UCSF is bounded.**

UCSF authors / JAMIA Open (A certified de-identification system for all clinical text documents) reports: The available paper summary discusses the clinical system’s development and precision concerns. Limit: No numeric result from this paper was independently verified for this guide. These signals add context, but they do not establish comparative performance.

- UCSF authors / JAMIA Open: The available paper summary discusses the clinical system’s development and precision concerns. (No numeric result from this paper was independently verified for this guide.)

## Caveats

**Keep the source boundary visible while you decide.**

The current cited cost statement is: BSD-3-Clause software; integration and compute separate. Privacy or processing control is described as: Privacy controls are not fully described in the cited sources. The principal product boundary is: Clinical-domain focus; general chat/email quality is not established. These are source-qualified statements, not an installation result, benchmark, security audit, or retention guarantee.

- The paper summary provides clinical context, but no independently verified numeric result supports a performance comparison here.
- Repository and published system versions may differ. No claimed compliance certification or accuracy number.

## First check

**Make the first adoption check specific to Philter UCSF.**

Run a labeled sample from the target corpus through detection and replacement, then inspect misses, over-redaction, and logs before sending real data through the pipeline.

## Claim Table

| Where | Claim | Confidence | Source | Caveat |
| --- | --- | --- | --- | --- |
| Workflow | Remove protected health information from clinical plain text. | High | UCSF / BCHSI (philter-repo) | Source-qualified product description; no independent performance test was run. |
| Setup | Local Python command-line package. | High | UCSF / BCHSI (philter-repo) | Source-qualified product description; no independent performance test was run. |
| Setup | Plain-text file or directory of clinical notes. | High | UCSF / BCHSI (philter-repo) | Source-qualified product description; no independent performance test was run. |
| Decision | BSD-3-Clause software; integration and compute separate. | High | UCSF / BCHSI (philter-repo) | Observed source wording; regional pricing, billing cadence, and future plan changes are not inferred. |
| Workflow | Configurable filtering pipeline with optional evaluation annotations. | High | UCSF / BCHSI (philter-repo) | Source-qualified product description; no independent performance test was run. |
| Workflow | Filter configuration and input/output paths. | High | UCSF / BCHSI (philter-repo) | Source-qualified product description; no independent performance test was run. |
| Caveats | Clinical-domain focus; general chat/email quality is not established. | Medium | UCSF / BCHSI (philter-repo) | This is a scope inference; published clinical results do not establish quality on general chat or email. |
| Review signal | UCSF authors / JAMIA Open reports: The available paper summary discusses the clinical system’s development and precision concerns. | Medium | UCSF authors / JAMIA Open (philter-paper) | No numeric result from this paper was independently verified for this guide. |

## Sources

- [Privacy Filter repository and CLI](https://github.com/openai/privacy-filter) - OpenAI; Official; accessed 2026-10-18
- [Presidio overview](https://presidio.dataprivacystack.org/) - Data Privacy Stack; Official; accessed 2026-10-18
- [Presidio Anonymizer](https://presidio.dataprivacystack.org/anonymizer/) - Data Privacy Stack; Official; accessed 2026-10-18
- [scrubadub README and license](https://github.com/LeapBeyond/scrubadub) - LeapBeyond; Official; accessed 2026-10-18
- [scrubadub stable documentation](https://scrubadub.readthedocs.io/en/stable/) - scrubadub maintainers; Official; accessed 2026-10-18
- [Pricing and deployment FAQ](https://www.getlimina.ai/en/pricing) - Limina; Official; accessed 2026-10-18
- [Philter UCSF README](https://github.com/BCHSI/philter-ucsf) - UCSF / BCHSI; Official; accessed 2026-10-18
- [GLiNER large-v2.1 vs Privacy Filter case study](https://github.com/gauravvij/pii-filter-model-eval) - Gaurav Vij / GitHub; Community; accessed 2026-10-18
- [A certified de-identification system for all clinical text documents](https://pmc.ncbi.nlm.nih.gov/articles/PMC10320112/) - UCSF authors / JAMIA Open; Review; accessed 2026-10-18

## Agent Retrieval Notes

- This Markdown is generated from the same structured guide package as the public HTML page.
- MCP tools return Markdown text by default to avoid high-token JSON envelopes.
- Agent site map: https://fieldguides.co/llms.txt