The 89f25c8 commit in ablation-tool/ablation adds GraphSession and a Doc-node-based vector search path to KnowledgeGraph. According to the repository description, Ablation is an autonomous reverse engineering system. This change is recorded in the latest commit dated 2026-10-25 and in the modified files.
GraphSession is a wrapper that pre-holds reverse engineering session information such as session_id, arch, and binary_path. When created via KnowledgeGraph.session(), there is no need to pass the session information again when recording results. Tool execution results are recorded with record(), findings with record_finding(), and the context manager is also supported. This replaces manual add_run() calls. Vector search embeds the text of Doc nodes with BAAI/bge-small-en-v1.5 and stores it in the embeddings table. If embeddings exist, documents are searched by the cosine similarity between the task description and the document embeddings, and Modules are ranked through DOCUMENTED_BY relationships. If embeddings are absent or the required libraries are unavailable, it falls back to workflow_stage ordering. An --embed option was added to the graph build script to run embeddings after bootstrapping and synchronization.
Embedding generation and semantic search require fastembed and numpy. The commit message records that it fixes a MEDIUM issue found while handling corrupted vector data with the stage-ordering fallback path, and reports that 70 KG tests and 1725 total tests pass. No comparison figures between the search results and the existing ordering method are given. Therefore, the effect of vector search is difficult to judge based on this commit's material alone.