Multimodal Mounting

IGraph supports mounting PRD documents and DB schemas as external resources into the code graph, establishing cross-modal associations based on semantic similarity.

What Is Multimodal Mounting?

Code doesn't exist in isolation. Requirements documents describe what code should do; database schemas describe the data structures code operates on. IGraph links these external resources to code symbols, so AI assistants can see relevant requirements and data models while understanding code.

Supported Resource Types

PRD Documents

Format	Extensions
Markdown	.md .markdown
Plain text	.txt
PDF	.pdf
Word	.docx

igraph mount prd docs/requirements.md
igraph mount prd specs/PRD.pdf

DB Schema

Format	Extensions
SQL DDL	.sql .ddl
JSON Schema	.json
Excel	.xlsx

igraph mount db schema/create_tables.sql
igraph mount db schema/er-model.json

Association Mechanism

When mounting, IGraph:

1. Parses resource content and extracts semantic fragments
2. Generates vectors for each fragment
3. Computes similarity against existing code symbols
4. Establishes association edges based on thresholds

Associations have two levels:

Level	Threshold	Description
Strong	≥ 0.85	High-confidence match; directly creates describes / reads edges
Weak	≥ 0.70	Optionally confirmed by LLM before creating edges

Configuration

{
  "multimodal": {
    "strongLinkThreshold": 0.85,
    "weakLinkThreshold": 0.7,
    "llmConfirmWeakLinks": false
  }
}

Field	Description
strongLinkThreshold	Strong association threshold (≥ this value = direct link)
weakLinkThreshold	Weak association threshold (between this and strong = optional confirmation)
llmConfirmWeakLinks	Whether to use LLM to confirm weak associations (improves accuracy but costs tokens)

Check Mount Status

igraph status

The output lists mounted resources and their association counts.

Resources Without Code Associations Are Still Retrievable

When a PRD describes a brand-new requirement (no corresponding code implementation yet), no resource_edges are created during mounting — because no code files reach the similarity threshold.

This doesn't affect retrieval: igraph_explore has a built-in independent resource retrieval channel that directly searches stored resource vectors and text, ensuring new PRDs aren't missed.

See Vector Retrieval — Independent Resource Retrieval.