[Enterprise](/content/enterprise/index.html)
Products
[Pricing](/content/pricing/index.html) [Docs](https://docs.potpie.ai/) [Blog](/content/blog/index.html)
[5k+](https://github.com/potpie-ai/potpie)
[Sign UP](https://app.potpie.ai/)
\\
\\
Specialists](/content/potpie-specialists/index.html) \\
\\
Forge](/content/potpie-forge/index.html) \\
\\
Recipes](/content/potpie-recipe/index.html) \\
\\
Trace](/content/potpie-trace/index.html)
Leveraging Codebase Knowledge Graphs for Agentic Code Generation
Dhiren Mathur
Oct 15, 2025
[TL/DR;](/content/blog/leveraging-codebase-knowledge-graphs-for-agentic-code-generation#header-0/index.html) [Alternative systems we experimented with:](/content/blog/leveraging-codebase-knowledge-graphs-for-agentic-code-generation#header-1/index.html) [System Architecture:](/content/blog/leveraging-codebase-knowledge-graphs-for-agentic-code-generation#header-2/index.html) [Flow for Generating Complete Code Context](/content/blog/leveraging-codebase-knowledge-graphs-for-agentic-code-generation#header-3/index.html) [Benefits](/content/blog/leveraging-codebase-knowledge-graphs-for-agentic-code-generation#header-4/index.html) [Conclusion](/content/blog/leveraging-codebase-knowledge-graphs-for-agentic-code-generation#header-5/index.html)
### TL/DR;
This blog explores how knowledge graphs can help language models generate code with better context by understanding the entire codebase. It discusses challenges like duplicating functionality, violating design patterns, and missing business requirements. The approach involves mapping function relationships, providing runtime code fetches, and delivering function-level explanations via a vector database for agents to query. This system optimizes code generation by providing detailed, dynamic context to avoid redundant, inconsistent code and ensure alignment with design and business logic across complex codebases.
An agent generating code with language models can face several challenges without sufficient context about the codebase:
- Duplicating functionality due to lack of awareness of existing functions and classes
- Violating design patterns and architectures, such as directly querying the database instead of using a repository pattern
- Incorrectly using new libraries and frameworks due to knowledge gap.
- Misaligning with business requirements and not fulfilling intended use cases due to lack of context
- Unable to integrate with existing data models and schemas, leading to duplicate fields and data inconsistencies
This article presents an approach for providing complete codebase context to a [**ReAct agent**](https://react-lm.github.io/) by:
1. Mapping function call relationship graphs in a graph database, detecting entry points into the server.
2. Fetching relevant code at runtime along with file paths
3. Generating function and entry point level explanations for the corresponding flows for every entry point.
4. Serving the explanations and metadata through a vector database that the agent can query with natural language
### Alternative systems we experimented with:
- Iterative mapping of function > class > file relationships such that the RAG system has the data about every function, class, file and can provide much more specific answers.
The cons associated with this approach were that:
\- It was slow
\- It was expensive
\- Caused us to hit rate limits on larger repositories
- For a use case which only involves understanding existing flows to derive context from them, moving to a system where we generate a flow level inference both for constituting functions of that flow and an overall intent of the flow provided the perfect trade-off between time, cost and quality.
### System Architecture:
Key components:
- Graph database (Neo4j) stores nodes representing functions and edges representing call relationships
- Postgres database:
- Endpoints table with API paths and identifiers
- Inferences table with intent extracted from explanations
- Explanations table with natural language descriptions of functions
- Pydantic table with data model definitions
- Relevant python services:
These services are part of the broader [**getmomentum/momentum-core**](https://github.com/getmomentum/momentum-core) project that is a code behaviour auditor which aims at analysing code behaviours at every git push.
- [**flow.py**](https://github.com/getmomentum/momentum-core/blob/main/server/knowledge_graph/flow.py): Traverses full project graph to generate explanations for each entry point upfront
- [**knowledge\_graph.py**](https://github.com/getmomentum/momentum-core/blob/main/server/knowledge_graph/knowledge_graph.py): Loads Postgres data into a vector database and provides method to query it.
- [**tools.py**](https://github.com/getmomentum/momentum-core/blob/main/server/test_agent/tools.py): Defines Langchain tools to fetch code, Pydantic definitions, query the knowledge graph. These tools will be used by code generation agents to gather pointed context.
### Flow for Generating Complete Code Context
1. In a processing step, [**flow.py**](http://flow.py/) is triggered for the entire project codebase right after parsing and endpoint detection is completed.
2. It captures all the functions, classes, entry points defined in the system.
3. For each entry point:
- It fetches the code for all functions in the flow from GitHub (we do not want to store code in our db.)
- It generates detailed natural language explanations of each function in the flow, capturing input, response, business logic, exceptions.
- It summarizes the overall intent of the endpoint from the function explanations. This is a single line statement of purpose of each API.
4. The endpoint paths, inferred intents, function explanations, and Pydantic models are loaded into the vector DB. We are using [**Weaviate**](https://weaviate.io/) as our vector DB of choice.
5. To load this data into Weaviate and query it, ideally, we would chunk, embed, set overlap, define query algorithm etc but for simplicity, here we use [**Embedchain**](https://github.com/embedchain/embedchain) to build the RAG tool.
6. When a user requests code generation for a specific entry point:
- The ReAct agent uses tools to:
- Directly fetch relevant code and Pydantic definitions for that entry point
- Query the vector DB with natural language to retrieve explanations and metadata for that endpoint and its functions
7. Equipped with granular code and high-level intent, the agent generates the requested code.
### Benefits
- Providing explanations for the full codebase upfront allows rapidly serving context for any entry point.
- Vector search enables dynamically retrieving only the most relevant explanations for the entry point of interest.
- Having context of other flows helps generate better results. Questions like "How can I create a document" can be answered with a reference to the POST /document/ API and file where it is present. This kind of queries asked by agents can help in tasks like writing a test for a DELETE /document/{id} endpoint that might need to insert a document first.
### Conclusion
- Mapping full codebases through a lens of entry points in vector databases provides a powerful tool for AI code generation agents.
- This approach equips language models with broad yet accessible context to generate code that respects the structure and intent of the full system.
- Techniques like this will be key to enabling agents to program by understanding entire codebases rather than just snippets in isolation.
" height="176.48689063250688px" id="t2zO9Quc7" stroke-dasharray="" stroke-linecap="round" stroke-linejoin="round" stroke-width="1.55" stroke="rgb(180, 209, 63)" transform="translate(0.773 0.758)" width="307.0499654602649px"/>)
" height="127.83155915492955px" id="F7uNl9dyd" stroke-dasharray="" stroke-linecap="round" stroke-linejoin="round" stroke-width="2.46" stroke="rgb(180, 209, 63)" transform="translate(8.021 100.549)" width="136.10429429577454px"/>)
" height="275.2348071827544px" id="tOCVqXAMo" stroke-dasharray="" stroke-linecap="round" stroke-linejoin="round" stroke-width="2.05" stroke="rgb(180, 209, 63)" transform="translate(2.023 1.876)" width="306.08189403973506px"/>)
" height="272.22086421478906px" id="PPsuL8gM9" stroke-dasharray="" stroke-linecap="round" stroke-linejoin="round" stroke-width="2.04" stroke="rgb(180, 209, 63)" transform="translate(2.016 1.798)" width="306.4324539473682px"/>)
" height="176.48689063250686px" id="YZ_qP17EP" stroke-dasharray="" stroke-linecap="round" stroke-linejoin="round" stroke-width="1.55" stroke="rgb(180, 209, 63)" transform="translate(0.773 0.758)" width="307.04996546026496px"/>)
" height="150.90560294117645px" id="jxii70vzV" stroke-dasharray="" stroke-linecap="butt" stroke-linejoin="miter" stroke-miterlimit="10" stroke-width="5.69" stroke="rgb(180, 209, 63)" transform="translate(24.37 4.791)" width="261.37546588235216px"/>)
" width="309px">)
" width="291.547144895145px">)
" width="279.949712674067px">)
[Try Potpie →](https://app.potpie.ai/) [Book a demo](https://cal.com/team/potpie/potpie-s-product-demo)
Recommended Blogs
\\
\\
Evals\\
\\
Evaluating AI Coding Agents in the Real World\\
\\
\\
\\
Kshitij Mishra\\
\\
Apr 9, 2026](/content/blog/evaluating-ai-coding-agents-in-the-real-world/index.html)
\\
\\
compliance\\
\\
Compliance Is Architecture: Why Your AI Agent Needs a Context Graph\\
\\
\\
\\
Yash Krishnan\\
\\
Apr 9, 2026](/content/blog/compliance-is-architecture-why-your-ai-agent-needs-a-context-graph/index.html)
\\
\\
Complaince\\
\\
Compliance as Architecture: Building Trustworthy AI from the Ground Up\\
\\
\\
\\
Yash Krishan\\
\\
Feb 2, 2026](/content/blog/compliance-as-architecture-building-trustworthy-ai-from-the-ground-up/index.html)
\\
\\
Product\\
\\
How Customer Onboarding Works at Potpie\\
\\
\\
\\
Kshitij Mishra\\
\\
Mar 18, 2026](/content/blog/how-onboarding-works-at-potpie/index.html)
\\
\\
Knowledge Graph\\
\\
Building a Distributed Knowledge Graph Pipeline\\
\\
\\
\\
Dhiren Mathur\\
\\
Mar 16, 2026](/content/blog/building-a-distributed-knowledge-graph-pipeline/index.html)
\\
\\
Evals\\
\\
Repository Intelligence: Building the Next Generation of Agent Evaluation Data\\
\\
\\
\\
Deeptendu\\
\\
Feb 27, 2026](/content/blog/the-agent-evaluation-gap/index.html)
\\
\\
Agents\\
\\
AI Memory is Broken: we need brains, not hard drives\\
\\
\\
\\
Deepesh Genani\\
\\
Feb 4, 2026](/content/blog/ai-memory-is-broken-we-need-brains-not-hard-drives/index.html)
\\
\\
Behind the scenes : Potpie SWE Bench Agent \\
\\
\\
\\
Nandan\\
\\
Jan 29, 2026](/content/blog/behind-the-scenes-potpie-swe-bench-agent/index.html)
\\
\\
Engineering Agents\\
\\
Agentic COSS 2025\\
\\
\\
\\
Aditi Kothari\\
\\
Mar 20, 2025](/content/blog/agentic-coss-2025/index.html)
\\
\\
LLMs\\
\\
Potpie supports Multi-LLMs\\
\\
\\
\\
Dhiren Mathur\\
\\
Feb 27, 2025](/content/blog/potpie-supports-multi-llms/index.html)
\\
\\
Open Source\\
\\
Open-source custom AI agents \\
\\
\\
\\
Aditi Kothari\\
\\
Mar 20, 2025](/content/blog/open-source-custom-ai-agents/index.html)
\\
\\
Product\\
\\
Release of New web access feature\\
\\
\\
\\
Ayush Thakur\\
\\
Feb 20, 2025](/content/blog/release-of-new-web-access-feature/index.html)
\\
\\
Product\\
\\
Deconstructing Potpie’s Debugging agent\\
\\
\\
\\
Dhiren Mathur\\
\\
Nov 15, 2025](/content/blog/deconstructing-potpie%E2%80%99s-debugging-agent/index.html)
Spec driven development for large scale engineering
Product
[Enterprise](/content/enterprise/index.html)
[Use-cases](/#product)
[Blog](/content/blog/index.html)
[Pricing](/#home-pricing)
[Potpie for VS Code](https://marketplace.visualstudio.com/items?itemName=PotpieAI.potpie-vscode-extension)
[Documentation](https://docs.potpie.ai/introduction)
Company
[Trust Center](https://trust.potpie.ai/)
[Press](https://www.forbes.com/sites/davidprosser/2026/02/23/how-poptie-plans-to-help-enterprises-smash-the-ai-agent-barrier/)
[Careers](https://wellfound.com/company/potpie-ai/jobs)
[Open Source](https://github.com/potpie-ai/potpie)
[Contact Us](/content/contact/index.html)
Legal
[Privacy Policy](/content/legal/privacy-policy/index.html)
[Terms](/content/legal/terms-of-service/index.html)
© 2026 Potpie. All rights reserved.
\[CODEBASE Q&A AGENT\]