Local-first AI debugging intelligence. Drop in a repo and an error log — AxionSys tells you exactly what's broken, why, and gives you the fix.

Overview

AxionSys is a local-first AI system that ingests codebases and error logs, performs hybrid retrieval across code chunks, applies LLM reranking for precision, and generates structured root-cause analysis with actionable code fixes.

Built and tested entirely on an AMD Radeon RX 9060 XT 16GB. No cloud. No NVIDIA. No compromise.

How It Works

Stage	Component
1	Log / Traceback / Error Message
2	Log Parser — traceback, log error, and free-form parsing with query extraction
3	Traceback-aware file pinning — files mentioned in the traceback are prioritised
4	Hybrid Retrieval — BM25 + FAISS across all code chunks
5	LLM Reranker — Qwen 3.5 9B
6	Dynamic Fusion Scoring — adaptive alpha/beta weighting
7	Root Cause Engine — Qwen 3.5 9B with causal chain reasoning
8	Fix Generator — unified diff output
9	Structured JSON Output

Key Capabilities

• Hybrid Retrieval — combines dense vector search (FAISS) with sparse keyword search (BM25) for higher recall and precision than either alone
• Traceback-aware Pinning — files explicitly mentioned in the traceback are pinned to the top of retrieval results with maximum score, ensuring ground-truth files always reach the LLM
• LLM Reranking — reranks retrieved chunks using Qwen 3.5 9B with dynamic alpha/beta fusion scoring that adapts based on agreement between hybrid and rerank scores
• Multi-format Log Parsing — accepts Python tracebacks (including Python 3.11+ ~~~/^^^ markers), .log files, and free-form error messages, extracting structured queries automatically
• Causal Chain Reasoning — traces bug propagation across multiple files with call chain visualization showing OK / BUG / CRASH status per node
• Unified Diff Output — generates git-style patches ready to apply directly to your codebase
• GitHub URL Ingestion — paste a GitHub URL directly; repos are cloned once and cached locally so repeated runs skip the download
• Session Management — each ingestion creates an isolated session with its own vector index, BM25 store, and results
• Fully Local — runs entirely on consumer AMD hardware with no cloud dependency

Demo Flow

1. Point AxionSys at a local repository path or GitHub URL
2. Paste an error log or traceback
3. Receive:
   • Root cause with causal chain reasoning
   • Call chain visualization across affected files
   • Concrete fix in unified diff format
   • Confidence scores for both diagnosis and fix
   • Top retrieved files ranked by hybrid, rerank, and final score
   • Full pipeline execution trace

Technical Stack

Component	Technology
API	FastAPI
Vector Search	FAISS
Keyword Search	BM25
Embeddings	Sentence Transformers
LLM Runtime	Ollama
Reasoning Model	Qwen 3.5 9B
Fast Tasks	Mistral 7B
Frontend	React + Vite
Hardware	AMD Radeon RX 9060 XT 16GB

Architecture

backend/
├── services/
│   ├── loader.py               # repo file loading
│   ├── chunker.py              # code chunking with overlap
│   ├── embedding.py            # sentence transformer embeddings
│   ├── vector_store.py         # FAISS index
│   ├── bm25_store.py           # BM25 index
│   ├── hybrid_retriever.py     # fused BM25 + FAISS retrieval
│   ├── reranker.py             # LLM reranking
│   ├── root_cause.py           # causal chain analysis
│   ├── fix_generator.py        # unified diff generation
│   ├── log_parser.py           # traceback + log + free-form parsing
│   └── llm.py                  # model router
pipelines/
├── ingest_repo.py              # repo ingestion pipeline
├── ingest_logs.py              # log ingestion pipeline
└── analyze.py                  # full analysis pipeline
frontend/
└── src/
    ├── App.jsx
    └── components/
        ├── IngestStep.jsx      # repo ingestion UI
        ├── AnalyzeStep.jsx     # log input + execution trace
        └── ResultsStep.jsx     # root cause, chain, fix, files, trace tabs

Running Locally

Prerequisites

# Install Ollama — https://ollama.com
ollama pull qwen3:9b
ollama pull mistral:7b

Install Dependencies

pip install -r requirements.txt

Start the API

uvicorn backend.main:app --reload

Start the Frontend

cd frontend
npm install
npm run dev

Run Analysis Directly

python -m pipelines.analyze

Sample Output

Error:   AttributeError: 'NoneType' object has no attribute 'get'

Root Cause:
  In cache.py, self.store is initialized to None in __init__,
  but get() calls self.store.get(key) without checking if store
  exists. When get() is called on an uninitialized Cache instance,
  it raises AttributeError.

Affected Files: cache.py, db.py
Confidence:     0.95

Call Chain:
  [OK]    db.py        __init__()
  [OK]    cache.py     __init__()
  [BUG]   cache.py     get()
  [CRASH] cache.py     get()

Fix (cache.py):
--- cache.py
+++ cache.py
@@ -1,7 +1,7 @@
 class Cache:
     def __init__(self):
-        self.store = None
+        self.store = {}

     def get(self, key):
         return self.store.get(key)

Fix Confidence: 0.90

Hardware

Built and daily-driven on AMD Radeon RX 9060 XT 16GB.
All inference runs locally via Ollama with ROCm backend.
No cloud API calls. No data leaves your machine.