| name | feishu-knowledge-ingest |
| description | batch ingest feishu folders and single attachments into report-first knowledge artifacts. use when chatgpt needs to read a feishu directory or a single shared file, classify files, extract text from supported attachments, and produce ingest-report.md, kb-items.jsonl, failed-items.jsonl, and memory.candidate.md without directly writing memory.md. best for feishu knowledge training, directory learning, policy/manual ingestion, and controlled docx/pdf parsing workflows. |
Feishu Knowledge Ingest
Use this skill to turn a Feishu folder or a single shared attachment into structured, reviewable knowledge outputs.
What this skill does
- Accept a Feishu folder link/token or a single shared attachment.
- Classify files into direct-read, download-and-parse, manual-review, or permission-blocked.
- Parse
.docx and .pdf in v0.1.
- Produce report-first outputs instead of writing
MEMORY.md directly.
- Preserve failures and uncertainty instead of guessing content.
Supported v0.1 scope
Inputs
- Feishu folder link or
folder_token
- Single shared attachment link or token
Parsing
Outputs
ingest-report.mdkb-items.jsonlfailed-items.jsonlMEMORY.candidate.md
Required behavior
- Distinguish Feishu native docs from uploaded attachments.
- Native docs:
doc, sheet, wiki, bitable
- Uploaded attachments:
.docx, .pdf, .pptx, other files
- Do not claim attachment content was learned unless text was actually extracted.
- Default to report-first. Do not write
MEMORY.md in v0.1.
- Record every failed file with a concrete reason.
- Prefer plain-text summaries over complex Feishu cards when reporting progress.
File routing rules
Direct-read
Treat these as direct-read only when the runtime has a reliable native-reader path:
Download-and-parse
Treat these as download-and-parse:
Manual-review
Route here when the file is out of scope or low-confidence in v0.1:
.pptx- images
- scans with no extractable text
- archives
- unusual file types
Permission-blocked
Route here when listing is possible but the file cannot be downloaded or read.
Standard workflow
- Resolve input type.
- Folder link/token -> enumerate files.
- Single file link/token -> build a one-file manifest.
- Create a batch record.
- Generate
batch_id.
- Record
started_at.
- Build a manifest.
- File name
- File token/link
- file type
- route decision
- Attempt extraction.
.docx -> use parsers/parse_docx.py.pdf -> use parsers/parse_pdf.py
- Produce structured outputs.
- success -> append to
kb-items.jsonl
- failure -> append to
failed-items.jsonl
- Summarize the batch.
- Write
ingest-report.md
- Write
MEMORY.candidate.md
- Finish the batch.
- Record
finished_at
- Never auto-write
MEMORY.md
Output contracts
kb-items.jsonl
Write one JSON object per successfully extracted knowledge item with at least:
batch_idsource_filesource_tokenfile_typetopiccontent_typesummaryextracted_atconfidence
failed-items.jsonl
Write one JSON object per failed or blocked file with at least:
batch_idsource_filesource_tokenfile_typefailure_reasonerror_detailsuggested_actionfailed_at
MEMORY.candidate.md
Include:
- batch header (
batch_id, started_at, finished_at, source_directory or source_file)
- grouped knowledge summaries
- source references
- confidence notes
- items needing review
ingest-report.md
Include:
- Batch summary
- Input scope
- File counts and routing counts
- Successful extraction summary
- Failures and risks
- Recommended next actions
Safety rules
- Never invent text that was not extracted.
- If parsing fails, say so plainly and log it.
- Treat filenames as hints only, never as proof of document contents.
- Keep sensitive data out of
MEMORY.candidate.md unless the workflow explicitly allows it.
Included files
run.py: minimal batch runner for local testingparsers/parse_docx.py: docx text extraction helperparsers/parse_pdf.py: pdf text extraction helperreferences/output_examples.md: sample output shapes and field guidanceREADME.md: setup and usage notes
