KodaSync
A Searchable Technical Memory for Developers
KodaSync is a developer knowledge workspace for code, technical notes, and imported documentation. Saved material can be grouped by project, embedded for semantic retrieval, and reused as chat context. A Monaco-powered Studio brings editing and focused AI code actions into the same workspace.
Project Links
Explore the developer knowledge workspace or inspect the semantic retrieval and AI pipeline behind it.
Saving knowledge is easy. Finding it again is the problem.
Code snippets, debugging notes, architecture decisions, and useful documentation accumulate across projects. Exact-word search helps when you remember what you saved, but a new problem often starts with an idea rather than a filename or phrase. I built KodaSync to give that knowledge a retrieval path and the context to make it useful again.
A note becomes more than stored text
FastEmbed turns a note’s title and content into a 384-dimensional vector. Stored alongside the note in PostgreSQL through pgvector, that representation makes similarity-based retrieval possible. Tags help organize the library; embeddings support a different way to find it.
Create a note
Capture code or technical knowledge.
Store content
Save text, language, and project association.
Generate tags
Enrich a new note in the background.
Create an embedding
Represent the content in 384 dimensions.
Persist the vector
Store it with the note using pgvector.
Retrieve by meaning
Compare it with an embedded query.
Searching for the idea, not the exact phrase
The semantic-search endpoint embeds a question and compares it with the authenticated user’s note vectors using cosine distance. The closest notes are returned as candidate context. This complements the Library’s keyword and tag filters when the wording of a question differs from the wording of a saved note.
The same question can mean something different in another project
An authentication question needs the context of the codebase you are working in. Selecting a project for chat restricts the candidate notes before semantic ranking, keeping unrelated project notes out of that retrieval step.
- Question
- Selected project
- Project notes
- Semantic ranking
- Relevant context
Chat starts with retrieval, not an empty prompt
Retrieval-augmented generation supplies relevant note text to an existing language model. KodaSync does not train a model on the library; it chooses context for each uncached question.
- 1.
Embed and scope
Embed the question, restrict notes to the authenticated user, and apply the selected project filter when present.
- 2.
Retrieve context
Rank by cosine distance and take up to three relevant notes, rather than sending the whole library.
- 3.
Build and stream
Combine note text, project context, the question, and conversation history, then stream the model’s response.
- 4.
Keep the conversation
Persist messages and cache the completed response for repeated user/project/question combinations.
One workspace, three ways to use the knowledge
Library
Store, pin, filter, and organize notes by project. Keep saved code and documentation editable and easy to revisit.
Studio
Work in Monaco and request focused AI-assisted fixes, explanations, security review, documentation, optimization, or test drafts for the editor’s content.
Chat
Ask questions with retrieved notes as context, optionally limiting retrieval to the project you are working on.
Documentation can enter the same knowledge system
URL import fetches a supplied page, removes common noise such as scripts, navigation, and styles, and extracts readable text. GitHub file links are converted to raw-file URLs, with language detection for common code extensions. Saving the import creates a note and embedding, so existing documentation can join the retrieval pipeline without being rewritten by hand.
Keeping the workspace responsive
Save first, enrich afterward
Ordinary note creation saves the record and dispatches tag generation and embedding creation through FastAPI BackgroundTasks. The creation response does not wait for that enrichment; URL imports use their own fetch-and-embed path.
Stream the answer as it arrives
FastAPI streams response chunks and the chat interface renders them incrementally. The interaction does not wait for the entire generated answer before showing text.
Cache repeated work, bound expensive requests
Redis caches semantic-search results, AI actions, and repeated chat responses. API rate limits bound expensive creation, import, and AI endpoints.
Testing the system around the model
GitHub Actions starts PostgreSQL with pgvector and Redis, then runs backend Pytest integration tests. The suite mocks AI and Redis calls while exercising note creation, retrieval, deletion, AI-action routes, and chat persistence. Generated answers can vary; the surrounding application behavior can still be checked independently.
Building memory around the model
The useful part of KodaSync is the pipeline around generation: deciding what to save, how to retrieve it, and which project’s context belongs in a conversation. Building that pipeline alongside an editable library, streamed chat, and testable API flows made saved technical knowledge reusable across the workspace.

