wiki-ingest

wiki-ingest: Source Ingestion into Knowledge Graph

You are a knowledge architect. Read the source, extract structured knowledge, and file it into the codescope knowledge graph. Every source should touch 5-15 graph nodes.

Ingestion Flow

1. Read the source

• File path: Read the file directly
• URL: Use WebFetch to retrieve content
• Image: Read the image, describe contents (OCR + visual analysis)

2. Extract structured knowledge

From the source, identify:

Entities (people, orgs, products, repos, technologies):

For each entity:
  - name (canonical form)
  - kind: person | org | product | repo | technology | standard | concept
  - description (1-2 sentences)
  - source: the file/URL this came from

Claims (factual assertions with confidence):

For each claim:
  - statement (one sentence, falsifiable)
  - confidence: high | medium | low
  - source: where this claim comes from
  - entities: which entities this claim is about

Concepts (ideas, patterns, frameworks):

For each concept:
  - name
  - description
  - related_concepts: other concepts this links to
  - related_code: function/class names in the codebase this maps to

3. File into the knowledge graph

For each extracted item, call the appropriate MCP tool:

knowledge_save(
  title: "OAuth2 Authorization Flow",
  content: "OAuth2 uses authorization codes...",
  kind: "concept",
  source_url: "https://...",
  tags: ["auth", "security"]
)

4. Cross-reference with code

After filing knowledge entities, link them to existing code:

knowledge_link(
  from_entity: "OAuth2 Authorization Flow",
  to_entity: "handle_auth_callback",
  relation: "implemented_by"
)

Use search_functions or semantic_search to find related code entities.

5. Check for contradictions

Search existing knowledge for claims that conflict with new ones:

knowledge_search(query: "OAuth2 token expiry")

If contradictions found, flag them:

knowledge_save(
  title: "Contradiction: token expiry time",
  content: "Source A says 1 hour, Source B says 24 hours",
  kind: "contradiction",
  tags: ["needs-resolution"]
)

6. Log the operation

memory_save(
  content: "Ingested [source]. Created N entities, M claims, L links.",
  kind: "operation"
)

Batch Mode

If user says "ingest all files in .raw/":

1. List all files in the directory
2. Process each sequentially
3. After all done, report summary: files processed, entities created, links made

Output Format

After ingestion, report:

## Ingested: [source name]

**Entities created:** 5
- OAuth2 (technology)
- Google Auth (product)
- handle_auth_callback (code link)
...

**Claims filed:** 3
**Cross-references:** 7 (4 to code, 3 to existing knowledge)
**Contradictions found:** 0

Knowledge graph updated.