KernelMind v2

Graph-aware repository intelligence engine for grounded code retrieval, workflow reconstruction, and repository reasoning.

Because semantic search alone is never enough.

---

What is KernelMind?

KernelMind is a retrieval + reasoning system built for understanding large repositories through:

• hybrid retrieval
• graph-aware expansion
• workflow reconstruction
• semantic reranking
• grounded answer generation
• repository-level reasoning traces

The system parses repositories into structured chunks, builds a call graph, retrieves semantically relevant regions, expands execution context using graph traversal, reranks candidates using a cross-encoder, and generates grounded responses from retrieved evidence.

KernelMind uses a hybrid architecture:

• grounded answer generation is currently handled through Sarvam AI
• retrieval, graph expansion, reranking, and evaluation run locally

Yes, my GPU has seen things.

---

Architecture Overview

Query
  │
  ▼
Query Classification
  │
  ▼
Hybrid Retrieval
(BM25 + Embeddings + RRF)
  │
  ▼
Query-Aware Seed Reranking
  │
  ▼
Graph Expansion
  │
  ├── Depth Decay
  ├── Propagation Scoring
  ├── Operation-Aware Traversal
  ├── Query Overlap Boosting
  └── Connectivity Weighting
  │
  ▼
Cross-Encoder Reranking
  │
  ▼
Reasoning Trace Construction
  │
  ▼
Context Building
  │
  ▼
Grounded Answer Generation

---

Retrieval Pipeline

KernelMind uses a multi-stage retrieval pipeline designed specifically for repositories.

Retrieval Stages

1. Embedding Retrieval

Semantic retrieval over chunk embeddings.

2. BM25 Retrieval

Exact lexical matching for:

• symbols
• identifiers
• filenames
• short tokens
• error strings

3. Reciprocal Rank Fusion (RRF)

Merges semantic and lexical retrieval.

4. Query-Aware Seed Ranking

Scores candidate seeds using:

• overlap
• chunk type
• propagation weighting
• retrieval confidence

5. Graph Expansion

Expands repository context using:

• forward traversal
• reverse traversal
• propagation scoring
• connectivity weighting
• operation-aware gating
• depth decay

6. Cross-Encoder Reranking

Final reranking pass using:

cross-encoder/ms-marco-MiniLM-L-6-v2

7. Grounded Generation

LLM generation constrained to retrieved repository evidence. Here, I use Sarvam AI, 105b model, which, by the way, is given for FREE. I know, it surprised me too.

---

Models

Embeddings

sentence-transformers/all-MiniLM-L6-v2

Cross Encoder

cross-encoder/ms-marco-MiniLM-L-6-v2

Generation Provider

Sarvam AI

KernelMind currently uses Sarvam AI for grounded answer generation.

Local Evaluation Model

Qwen 7B via Ollama

Local Runtime Components

The retrieval stack, graph expansion, reranking pipeline, vector search, and evaluation infrastructure run locally.

Current local components include:

FAISS
BM25
Cross-Encoder Reranking
Qwen 7B via Ollama
Sarvam AI

---

Storage + Retrieval Infrastructure

Vector Database

FAISS

Additional Retrieval Systems

• BM25
• graph propagation
• reranking layers
• workflow reconstruction
• reasoning traces

---

Supported Languages

Currently supported:

• Python
• Eventually working to add other languages - feel free to help!

Additional languages can be added through parser extensions.

---

Query Modes

KernelMind routes queries into different execution modes.

Workflow Mode - Deep graph traversal and execution reconstruction.

Symbol Lookup - Precision-oriented retrieval with minimal expansion.

Architecture Mode - Broader repository exploration.

General QA - Balanced retrieval fallback.

---

TUI Interface

KernelMind includes a Textual-based terminal interface featuring:

• repository querying
• retrieval observability
• graph traces
• ranked chunk inspection
• streamed grounded answers
• workflow visualization

---

RAGAS Evaluation

KernelMind includes evaluation support using:

RAGAS

with local Ollama-hosted Qwen evaluation.

Results

Configuration	Faithfulness	Answer Relevancy	Context Precision	Context Recall
Graph Expansion Enabled	0.6080	0.7697	0.5962	0.5357
Graph Expansion Relaxed	0.6356	0.8478	0.4732	0.5714

---

Interpreting The Results

The evaluation surfaced an important retrieval tradeoff:

• stricter graph traversal improved precision
• relaxed traversal improved workflow continuity and answer relevancy
• broader graph expansion improved recall
• excessive traversal suppression reduced execution-chain visibility

This is currently one of the most interesting active areas of experimentation in the project.

The system is effectively balancing:

precision ↔ recall
semantic locality ↔ workflow continuity

which is a surprisingly deep rabbit hole once graph propagation enters the picture.

---

Current Features

• Hybrid retrieval
• FAISS vector search
• BM25 lexical retrieval
• Graph-aware expansion
• Propagation scoring
• Cross-encoder reranking
• Query-aware traversal
• Workflow reconstruction
• Reasoning traces
• Grounded generation
• Persistent repository runtimes
• Textual-based TUI
• Local-first execution - well, mostly!

---

Example CLI Flow

python -m app.tui

Query → Retrieval → Expansion → Reranking → Grounded Response

---

Why KernelMind Exists

Most repository assistants:

• retrieve disconnected chunks
• ignore execution structure
• lose workflow continuity
• hallucinate relationships
• flatten repositories into embeddings

KernelMind attempts to preserve:

• execution topology
• call relationships
• workflow causality
• repository structure
• retrieval transparency

while still remaining practical to run locally.

---

Contributions

Contributions are welcome.

Especially if you enjoy:

• retrieval systems
• graph reasoning
• semantic search
• code intelligence
• information retrieval
• systems engineering
• debugging graph propagation at 2 AM

Open issues, ideas, experiments, benchmarks, weird retrieval failures, and improvements are all appreciated.

---