Embedded Agentic RAG Stack for multi-repo code intelligence with FastAPI, MCP, and knowledge graph
Alethe-RAG provides 11 tools with consistent naming (verb_noun pattern: ask_about_*, get_*, find_*, list_*, run_*) and descriptions present for all tools. However, descriptions are often generic or lack actionable context for LLM selection. Parameter schemas are visible and typed, but descriptions for parameters are minimal or missing in several tools. Output schemas are not documented, responses are returned as JSON strings without declared structure. Error handling is absent; no guidance on recovery or retry logic. The server uses HTTP transport (FastMCP) which is current, but lacks tool annotations (readOnlyHint, destructiveHint) and structured output documentation.
Ask a question about a specific class/module in the project.
Ask a natural-language question about the indexed source code. Uses hybrid RAG: embeds the question, runs vector k-NN search against code nodes, then enriches results with graph context (neighbors, relationships). Good for: 'How does the payment flow work?', 'Which classes depend on OrderService?', 'Find authentication logic'.
Find the shortest dependency path between two nodes.
Get the dependency subgraph for a class.
Returns knowledge graph statistics and breakdowns.
Get a complete map of how repositories connect through infrastructure. Shows which repos expose endpoints, share databases, use the same queues, emit/handle the same events. This is the high-level architectural view of the entire system.
Output schemas not documented. Tools return JSON strings without declared field structure. LLMs cannot plan downstream calls or extract specific fields reliably.
Parameter descriptions missing or minimal. 'class_name', 'source_name', 'target_name', 'cypher', 'node_name' lack guidance on format, constraints, or examples. LLMs cannot validate input before calling.
No error handling or recovery guidance. Tools do not document what happens on invalid input, missing nodes, or query failures. No actionable error messages to guide LLM retry logic.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Get full details of a node by name.
List cross-repository infrastructure bridges. Bridges connect code in one repo to code in another through shared infrastructure: HTTP endpoints, databases, queues, events, gRPC. Also surfaces implicit cross-repo AST edges (extends, implements, depends_on) where the source and target belong to different repos.
List all detected infrastructure nodes: HTTP endpoints, databases, message queues, domain events, background jobs, gRPC services. These are virtual nodes that represent shared resources between repos. Each shows which repos expose/consume it.
List nodes in the knowledge graph with optional filters.
Execute a raw Cypher query on the GraphQLite knowledge graph.
Tool annotations missing. No readOnlyHint, destructiveHint, or idempotentHint declared. LLMs cannot determine which tools are safe to retry or have side effects.
Pagination not implemented for list tools. 'list_nodes' and 'list_bridges' accept 'limit' but no offset/cursor. Large result sets will blow context window without proper pagination.