Add AI Cordon Picket integration - #576
Conversation
|
@mihail-gribov is attempting to deploy a commit to the deepset Team on Vercel. A member of the Team first needs to authorize it. |
There was a problem hiding this comment.
🟡 Changes recommended
The integration page sets toc: true but is missing the repository’s standard “Table of Contents” block, making it inconsistent with the documented integration template.
Once you've addressed the issues Copilot identified, you can request another Copilot review.
Pull request overview
- Adds a new integration card for the
aicordon-haystackpackage, documenting two Haystack components intended to detect prompt injection at ingest-time and pre-generation time.
Changes:
- Adds installation instructions and usage examples for
PromptInjectionFilter(ingest/documents) andPromptInjectionGuard(generator/messages). - Documents routing behavior (
blockedsocket) and includes benchmark/latency figures plus license details.
File summaries
| File | Description |
|---|---|
integrations/aicordon-haystack.md |
New integration documentation page for AI Cordon Picket components, including usage patterns, measured results, and license. |
Review details
- Files reviewed: 1/1 changed files
- Comments generated: 1
- Review effort level: Lite
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
| Check what an LLM is given for prompt injection - in both places it can arrive. | ||
|
|
| socials: | ||
| github: mihail-gribov | ||
| pypi: https://pypi.org/project/aicordon-haystack | ||
| repo: https://github.com/AICordon/aicordon |
There was a problem hiding this comment.
| repo: https://github.com/AICordon/aicordon | |
| repo: https://github.com/AICordon/aicordon/tree/main/integrations/haystack |
We typically link the integration itself, not the main repository.
| | `PromptInjectionFilter` | **material**: documents at ingest, before they are chunked and embedded | Picket's `ipi` rules | | ||
| | `PromptInjectionGuard` | **the request**: the turn the model is about to answer | Picket's `dpi` rules | |
There was a problem hiding this comment.
Could you please define the Picket's rules? It might be unclear.
| The check is a rule, not a model: no GPU, no network, no key, a few hundred kilobytes of base, and a | ||
| fraction of a millisecond per turn on one core. |
There was a problem hiding this comment.
Could we elaborate on the algorithm used by aicordon? Since it is supposed to intercept the communication with the LLM, that might be crucial for our users to understand if that's safe enough.
Two Haystack components from
aicordon-haystack, one for each place a pipeline holds text a model is about to read:PromptInjectionFilter(preprocessors) checks documents at ingest, before chunking and embedding;PromptInjectionGuard(validators) checks the message list on its way into the generator, and routes a flagged exchange to ablockedsocket instead of calling the model.The check is a local rule base - no GPU, no network, no API key, no dependencies beyond
aicordon. Measured on the pipeline rather than on the detector, with the numbers and the harness in the package README.Card follows the schema in this repo's README: one card per package,
pypiandrepoboth resolve.