Mikk - A Deterministic Context Engine for AI-Assisted Development

[Ansh-dhanani]

Mikk 2.0 — A Deterministic Context Engine for AI-Assisted Development

Modern AI coding tools rely heavily on retrieval pipelines:

• indexing codebases

• generating embeddings

• returning “similar” snippets

This approach treats code as text. In reality, code is a structured system of relationships.

---

What Mikk Does Differently

Mikk replaces retrieval with deterministic context computation.

Code → AST → Symbol Graph → Scoring → Decision → Optimized Context

Instead of returning approximate matches, Mikk computes the exact set of code required for a given task.

---

System Architecture (Programmatic Overview)

1. Parsing (AST Generation)

• Uses Tree-sitter to convert source files into Abstract Syntax Trees

• Language-agnostic and structurally reliable

2. Symbol Extraction

• Converts AST nodes into symbols:

  • functions

  • classes

  • variables

• Each symbol becomes a node in the system

3. Dependency Resolution

• Resolves calls, imports, and references

• Builds explicit relationships between symbols

4. Graph Construction

• Directed, symbol-level dependency graph

• Nodes = symbols

• Edges = relationships (CALLS, IMPORTS, USES)

5. Query Processing and Traversal

• Task → keyword + semantic seeds

• Breadth-first traversal expands relevant dependencies

• Ensures inclusion of all required execution paths

---

6. Scoring Engine

Each node is ranked using:

Score =
  Relevance (task alignment)
+ Proximity (graph distance)
+ Centrality (structural importance)
+ Edge Confidence (resolution reliability)
+ Semantic Similarity (optional)

---

7. Token Budget Optimization

Select nodes by (score / token_size)

• Greedy knapsack approximation

• Maximizes information density within LLM constraints

---

8. Deterministic Compression

• Removes comments and non-essential tokens

• Collapses structural blocks

• Retains signatures + core logic

No generative summarization → fully deterministic.

---

Why This Matters

Traditional systems:

"What code looks relevant?"

Mikk:

"What code is required to safely perform this task?"

---

Practical Impact in VS Code

1. Accurate AI Assistance

• Complete dependency chains

• Reduced hallucinations

2. Safe Refactoring

• Full impact awareness

• No missed dependencies

3. Faster Navigation

• Focused execution paths

• No manual tracing

4. Productivity Gains

• Less search time

• More task execution

---

Example

Task:

"Add 2FA to login flow"

Mikk computes:

• login handler

• authentication middleware

• token validation logic

• database layer

---

Comparison with Existing Approaches

Feature | Retrieval Systems | Graph Tools | Mikk -- | -- | -- | -- Core Approach | Similarity search | Graph querying | Context decision Dependency Awareness | Partial | Strong | Complete Token Optimization | Limited | Limited | Built-in Determinism | Low | Medium | High Task Alignment | Indirect | Indirect | Direct

---

Key Insight

Search retrieves possibilities.
Decision constructs correctness.

---

Production Considerations

• Deterministic outputs (repeatable)

• Language-agnostic parsing

• Scalable graph architecture

• IDE integration ready

With caching + incremental updates, Mikk becomes a reliable infrastructure layer for AI-assisted development.

---

Conclusion

Mikk shifts the paradigm from retrieval to decision.

By treating code as a structured system, it enables:

• higher correctness

• better efficiency

• production-grade reliability

Especially within environments like VS Code.

[Ansh-dhanani]

https://github.com/Ansh-dhanani/mikk/