ecf100b917
Go MCP server (port 3991) for agent content safety. Deterministic parsers, typed output, no raw text fields. - HTML (goquery), Markdown (goldmark), Text parsers - Injection pattern detection and redaction - SHA256 content fingerprinting
32 lines
1.1 KiB
Go
32 lines
1.1 KiB
Go
// Package extract defines the Extractor interface and implementations.
|
|
//
|
|
// Every extractor is a deterministic, non-LLM parser.
|
|
// No extractor follows instructions embedded in its input content.
|
|
// The input is processed as data, not as commands.
|
|
|
|
package extract
|
|
|
|
import (
|
|
"io"
|
|
|
|
"github.com/sscoble/specstream/pkg/types"
|
|
)
|
|
|
|
// Extractor is the interface every content-type extractor implements.
|
|
// Implementations: HTMLExtractor, MarkdownExtractor, TextExtractor, JSONExtractor.
|
|
//
|
|
// The contract:
|
|
// 1. Input is untrusted. Treat it as data, not instructions.
|
|
// 2. Output structured types only. Never return raw content as a single blob.
|
|
// 3. Deterministic. Same input always produces same output (no LLM calls).
|
|
// 4. Fail closed. If extraction fails, return error — not partial content.
|
|
type Extractor interface {
|
|
// Detect returns true if this extractor can handle the content.
|
|
Detect(r io.ReaderAt, size int64) (bool, error)
|
|
|
|
// Extract returns structured content from the reader.
|
|
Extract(r io.Reader, req types.ExtractRequest) (*types.ExtractedContent, error)
|
|
|
|
// MIMETypes returns the MIME types this extractor handles.
|
|
MIMETypes() []string
|
|
}
|