graphiffy-abap

Knowledge-graph builder for ABAP / SAP codebases. Hard fork of safishamsi/graphify with first-class ABAP support backed by abaplint.

Point it at a folder of .abap files (raw dumps, abapGit-exported repos, or a cloned ABAP GitHub repo) and it produces a queryable knowledge graph of classes, interfaces, methods, function modules, forms, includes, table accesses, exception flow, and rationale comments — all with cross-file resolution.

What it does

Three-pass pipeline inherited from upstream graphify:

1. AST extraction (per file, deterministic, no LLM). For ABAP, this calls abaplint via a Node subprocess and walks its parsed statement tree to emit graph nodes/edges. Other languages still go through tree-sitter (Python, JS, Go, Java, etc.).
2. Cross-file resolution. Definition-site nodes and reference-site nodes share canonical IDs (clas:zcl_foo, fm:z_post_invoice, method:zcl_foo::process, ddic:vbak), so calls/inheritance/interface implementations stitch automatically across files.
3. LLM semantic pass (optional, --mode deep). Claude reads docs, PDFs, images, and rationale comments to add INFERRED edges and design rationale.

Output:

graphiffy-abap-out/
├── graph.json       canonical knowledge graph
├── graph.html       interactive graph (vis.js, open in any browser)
├── GRAPH_REPORT.md  god nodes, communities, surprising connections, suggested questions
└── cache/           SHA256-keyed per-file cache; re-runs only re-process changed files

What ABAP support extracts

For every .abap file, the pipeline captures:

Edge types:

• inherits_from, implements_interface, defines, includes
• calls_form (PERFORM)
• calls_function_module (CALL FUNCTION 'Z_FOO')
• calls_method (obj->meth, cls=>meth, legacy CALL METHOD)
• instantiates (NEW zcl_foo( ) / CREATE OBJECT)
• accesses_table (Open SQL)
• raises_exception, rationale_for
• *_dynamic (PERFORM (lv_x), CALL FUNCTION (lv_x)) — flagged AMBIGUOUS

Every edge carries confidence ∈ {EXTRACTED, INFERRED, AMBIGUOUS} and a confidence_score so the downstream report can be honest about what's solid vs. guessed.

Quick start

Prerequisites: Python 3.10+, Node.js 18+, an ABAP corpus.

# Install graphiffy-abap from a local clone (PyPI publication TBD)
git clone https://github.com/NabheetCloud/graphiffy-abap.git
cd graphiffy-abap
pip install -e .

# In the project you want to analyse, install abaplint
cd /path/to/your/abap-project
npm install @abaplint/core

# Run
graphiffy-abap .

Output lands in graphiffy-abap-out/ next to your source.

Worked example: abap2UI5

A real run on the public abap2UI5/abap2UI5 codebase produced a 2,700-node / 2,870-edge graph with 1,338 cross-file edges. Top god nodes (degree-ranked) were exactly what an SAP engineer would predict: CX_STATIC_CHECK, Z2UI5_CL_UTIL_EXT, LTCL_UNIT_TEST, CL_ABAP_UNIT_ASSERT::ASSERT_EQUALS, Z2UI5_IF_AJSON.

Full validation report: worked/abap2ui5/review.md

File detection

graphiffy-abap recognises:

• Plain .abap files
• abapGit conventions: *.clas.abap, *.prog.abap, *.intf.abap, *.fugr/*.abap (function-group includes)
• Object type detected from filename (so a *.clas.abap is correctly typed as a class even before the source is parsed)

Other graphify file types (.md, .pdf, images, code in 25 other languages) are still supported — drop ABAP, design docs, and screenshots into one folder and graphiffy-abap will produce one unified graph.

Roadmap

Phase 1 (done). ABAP language plugin via abaplint. Programs, classes, interfaces, methods, function modules, forms, includes, calls, table references, exceptions, rationale.

Phase 2 (done). DDIC XML parsing (tables, structures, data elements, domains, search helps), CDS view DDL, AMDP method bodies, BAdI / enhancement framework awareness, SAP-aware Pass-3 LLM prompts.

Phase 3 (WIP). Live SAP system extraction via ADT REST / RFC, transport-aware incremental updates, on-prem LLM option, SAP-specific MCP server.

Relationship to upstream graphify

graphiffy-abap is a hard fork of safishamsi/graphify v5 (commit 770d7f5). The MIT license is preserved. ABAP support adds an external Node.js dependency (abaplint) and SAP-specific extraction logic that aren't generally useful to upstream's audience, hence the fork. See NOTICE for full attribution.

All non-ABAP languages still work exactly as upstream — graphiffy-abap is a strict superset of graphify's language support.

License

MIT — see LICENSE.