Project walkthrough · By Jamil Fayad
Personal RAG and AI-agent toolkit
Personal engineering toolkit · Source review: 20 October 2026
This personal Python toolkit explores using retrieved technical knowledge to support development agents. The walkthrough is based on inspection of the local source files; it is not a report of a live business deployment.
The problem and scope
A development assistant needs relevant context and a defined task. The toolkit separates document processing, embeddings, vector storage and querying, alongside separate agent orchestration, so those responsibilities can be examined independently.
My contribution
- Created document-processing code for Markdown, JSON and text, with configurable chunking and source metadata.
- Implemented persistent ChromaDB collections, local and optional cloud embedding adapters, and a query engine returning document IDs, source paths and scores.
- Added multi-domain search and context formatting, plus a separate orchestrator connecting development agents and collecting generated artifacts.
Reviewed source · 29 October 2026
What the retrieval code actually does
For an internal knowledge assistant, finding a passage and knowing where it came from are separate responsibilities. These excerpts show how my toolkit requests relevant passages and keeps source metadata with each result.
- Search scope: select a collection, request a limited number of results and optionally filter metadata. Metadata filtering is not an access-control system; a business deployment needs separate authorization checks.
- Traceability: carry the retrieved text, document ID and source-file metadata into the result. Source metadata does not prove an answer is correct or that generated citations match the evidence.
Inspect the Python excerpts
From rag_system/core/query_engine.py. Two excerpts, with indentation normalized and intervening code omitted; this is not a complete runnable example.
query_embedding = self.embedding_generator.generate_embeddings([query])[0]
# Query vector store
results = self.vector_store.query(
collection_name=domain,
query_embeddings=[query_embedding],
n_results=top_k,
where=metadata_filter,
)search_results.append(
SearchResult(
content=results["documents"][0][i],
score=similarity,
metadata=metadata,
document_id=results["ids"][0][i],
source_file=metadata.get("file_path", "unknown"),
)
)Source inspection, not a runtime demonstration. Ingestion, provider calls and end-to-end retrieval were not run for this review. Accuracy, permission enforcement and answer quality still need testing for an agreed business workflow.
Download the reviewed excerpts and limitations (.txt)How the parts connect
- 01Technical documents split into chunks with metadata
- 02Embeddings stored in domain-specific ChromaDB collections
- 03Query engine retrieves source-tagged context
- 04Agent output and generated artifacts for developer review
Why this is relevant to business software
The relevant pattern is connecting a maintained knowledge collection to a bounded assistant. The implementation provides material for discussing retrieval and orchestration choices; it does not establish that this toolkit is ready for sensitive business information. Retrieval and action are separate concerns here: the query engine returning source-tagged context is evidence of retrieval, not of safe tool-action execution. The separate orchestrator that connects development agents and collects generated artifacts is the closer analogue to permissioned actions, and it carries its own authorization and sandboxing checks below, independent of whether retrieval is involved.
What a business deployment would need
- Reproduce the environment and run ingestion, retrieval and provider calls end to end.
- Evaluate known-answer, no-answer and conflicting-source questions; check source attribution.
- Validate indexing across repeated document additions, deletion and updates.
- Enforce authorization before retrieval and model calls; test prompt injection and information leakage.
- Review generated code, sandbox tool access and test failures, costs and latency before release.
Project context
This is a personal engineering project, not a client case study. Source inspection supports the architecture described here; the toolkit was not executed as part of this review. No benchmark, accuracy improvement, production deployment or security assurance is claimed. Source-file metadata is present in the query engine; its correctness and usefulness in generated answers still need testing. The separate orchestrator imports agent modules from an external package path, so a reproducible environment is also needed.
Have a related workflow?
Describe your current tools, users and the work you want to improve. Smaller projects are welcome; we can agree a useful scope before quoting.