Bibliography
What SourcePrep was built on. Notes on the papers, repositories, essays, and standards that shaped the project.
Bibliography
A working list of the papers, repositories, essays, and standards SourcePrep draws on. Each entry includes a one-line note on how it shaped the project — and what we changed when we disagreed.
Why context engineering matters more than raw context size — and what changes when language models meet long, noisy windows.
Liu et al. show that language models attend to the start and end of a long context far more than the middle. The finding sets the ceiling for any retrieval system that pads its context naïvely. SourcePrep ranks results by relevance and assembles them so the highest-scoring chunks bracket the prompt, never bury it.
Han et al. compare flat-vector RAG against graph-augmented RAG across reasoning-heavy benchmarks and find that local community search wins on multi-hop questions. SourcePrep’s prep_search follows the same logic: vector hits seed the query, then a trace-graph hop expands the neighborhood before the final assembly.
Anthropic’s post argued that prepending a few lines of file-level context to each chunk before embedding reduced retrieval failures by 49% in their tests. SourcePrep’s semantic chunker now does exactly this — every chunk carries a synopsis prefix derived from its enclosing module so the embedding sees the same neighborhood the model will reason over.
Chen et al. · EMNLP 2025 Findings · 2025
Reasoning degrades even with 100% retrieval accuracy — evidence for retrieve-then-solve over context padding.
Chroma Research · trychroma.com · 2025
18-LLM evaluation showing distractor similarity compounds context degradation. Sets default min-score thresholds.
Databricks Research · databricks.com · 2024
Identifies the 4K–32K token RAG saturation point that justifies SourcePrep’s 6K–8K conservative defaults.
arXiv 2025 · 2025
Comprehensive map of code-RAG techniques. Used to position SourcePrep inside the repository-level RAG quadrant.
arXiv 2025 · 2025
Umbrella framing for the broader context-engineering discipline.
GitHub
MCP server reporting ~40% token reduction via vector search — external corroboration of SourcePrep’s retrieval-first approach.
Why SourcePrep’s context assembler ladders code from full source down to one-line signatures, and the research that makes signature-only context defensible.
Stingy Context demonstrates that hierarchical level-of-detail extraction can compress code by 18:1 with negligible quality loss for auto-coding tasks. SourcePrep’s context assembler implements the same ladder: full source for the focal file, compressed forms for callees, and one-line signatures for everything else.
HCP measures the cost of dropping function bodies in favor of signatures across three pruning depths. The result — ~90% retention of completion quality — is the empirical backbone for SourcePrep’s LOD 2 default. SourcePrep promotes bodies back to LOD 0 only where the search score crosses a threshold.
Aider’s repo-map prunes a project into one-line signatures and ranks the visible set per turn. It is the public proof that LOD 4 holds up in production across the open-source agentic-coding community. SourcePrep’s context assembler implements a similar selection step on top of the trace graph rather than a flat AST extract.
LLMLingua-2 uses a small BERT classifier to drop the lowest-information tokens from a prompt without losing meaning. SourcePrep runs it over Markdown and docstrings while letting code chunks flow through the structural LOD ladder — two compressors, one assembly. The repo hosts both versions; v2 is what SourcePrep adopts.
Wu et al. · ICML 2024 · 2024
Validates score-gated retrieval — knowing when *not* to fetch context improves accuracy.
Liu et al. · ASE 2024 · 2024
Baseline for graph-vs-embedding retrieval comparisons.
Phan et al. · arXiv 2024 · 2024
The Search→Expand→Refine pipeline maps 1:1 onto SourcePrep’s search → trace expansion → LOD assembly.
Liu et al. · arXiv 2024 · 2024
Static-analysis-at-prompting pattern; mirrors SourcePrep’s use of trace-graph import edges to drive dependency-aware retrieval.
Guo et al. · arXiv 2026 · 2026
Flagged as a Phase-2 enhancement — inline callees/callers on top of existing LOD results.
Zhang et al. · ICLR 2024 · 2024
Model-internal KV compression — explicitly complementary to SourcePrep’s pre-prompt compression layer.
Pan et al. · ACL 2024 · 2024
BERT-classifier token pruning. Adopted as the language/docs compressor.
Hai et al. · NAACL Findings 2025 · 2025
Empirical evidence that signatures + docstrings are the highest-ROI context type.
GitHub
Production tree-sitter compression at ~70% reduction — evidence that LOD extraction is practical at scale.
GitHub
Evaluated as an off-the-shelf compressor; rejected due to 7B-model dependency incompatible with local-first architecture.
arXiv 2025 · 2025
Bigraph retrieval reference — supports the case for graph-structured code representation.
Why chunking on AST boundaries beats character splits for code, and how structural awareness changes retrieval quality.
Garry Tan’s gbrain pairs Savitzky–Golay smoothing for semantic boundary detection with reciprocal-rank-fusion across multiple query expansions. Reading it shaped SourcePrep’s current chunker: smooth the similarity curve, cut on local minima, fuse vector and keyword hits with RRF.
cAST shows that chunking on AST boundaries produces meaningfully better embeddings than fixed-window splits, particularly for languages with strong nesting structure. SourcePrep’s tree-sitter chunker is grounded in this finding — a chunk never splits mid-function and the chunk header carries the full enclosing path.
Edge et al. layer entity extraction, community detection, and per-community summarization to make a knowledge graph queryable as a hierarchy. SourcePrep’s atlas does the same trick on code: directories and modules become communities, each with a generated synopsis that the assembler can hand to the model in place of the underlying files.
Jina AI · Jina blog · 2024
Evaluated for long-context code files; kept as exploratory.
Khattab & Zaharia · SIGIR 2020 · 2020
Token-level late interaction — reference paradigm for retrieval-quality upgrades.
Cormack, Clarke, Buettcher · SIGIR 2009 · 2009
RRF (k=60) is the hybrid-fusion strategy SourcePrep uses to combine semantic and keyword search.
Shahul Es et al. · EACL 2024 · 2024
Context-precision and context-recall metrics used to evaluate chunking-strategy experiments.
Savitzky & Golay · Analytical Chemistry · 1964
The smoothing filter behind semantic-boundary detection on similarity curves.
arXiv 2024 · 2024
Multi-pass LLM augmentation pattern that informs SourcePrep’s per-node enrichment pipeline.
Why SourcePrep treats concepts as first-class artifacts, where the protocol surface comes from, and the older work that grounds the system in something deeper than recent papers.
Guo et al. show that pre-training code models on data-flow graphs (not just token streams) measurably improves downstream code understanding. SourcePrep’s trace index encodes the same intuition by carrying control- and data-flow edges alongside symbol references, so retrievers can hop on dependency, not just on text similarity.
Nonaka & Takeuchi’s SECI model frames organizational knowledge as a four-step cycle: socialize, externalize, combine, internalize. SourcePrep’s concepts feature is a literal externalization tool — the tacit "we don’t do it that way" assumptions in a team’s head become typed, anchored, testable artifacts that downstream agents can read.
MCP is the protocol surface SourcePrep ships its primary interface on. Every prep_* tool, the resources system, and the per-client context budgets are MCP-shaped from the ground up. Without this spec there is no SourcePrep MCP server, and the page SourcePrep advertises to any agent in any IDE would not exist.
Ferrante, Ottenstein, Warren · ACM TOPLAS · 1987
Classical PDG reference; grounds SourcePrep’s combined control-flow + data-flow trace graph.
Traag, Waltman, van Eck · Scientific Reports · 2019
Theoretical backing for community-detection-driven concept clustering.
Ganter & Wille · Springer · 1999
Mathematical justification for lattice-based concept organization.
Ruven Brooks · IJMMS · 1983
Top-down program comprehension theory — the cognitive basis for hypothesis-driven retrieval.
Nancy Pennington · Cognitive Psychology · 1987
Bottom-up comprehension counterpart to Brooks; SourcePrep’s structural trace index supports this mode.
George A. Miller · Psychological Review · 1956
Cognitive-load rationale for concept clustering at human-readable cardinality.
Michael Nygard · cognitect.com · 2011
ADR template convention; SourcePrep concepts extend ADRs beyond per-node decisions.
ISWC · 2024
Establishes SOTA for automated concept extraction; informed SourcePrep’s hybrid embedding+LLM concept-discovery pipeline.
arXiv 2021 · 2021
Researched for requirements↔code linking; rejected as too heavy for the SourcePrep architecture but kept as a baseline.
NASA SWE Handbook
Grounds SourcePrep’s curated traceability framework in an established engineering standard.
agentclientprotocol.com
Zed-backed standard — SourcePrep’s multi-editor integration target.
Google / Linux Foundation · a2a-protocol.org · 2025
Identified as a future Layer 4 target for cross-agent discovery.
OASIS · 2020
SARIF-in / SARIF-out enrichment is a shipped prep_audit capability.
ocsf.io
Alternative audit-export format; AWS/Splunk-backed.
agents.md
Emerging convention for agent-facing context files — SourcePrep auto-generates AGENTS.md via rules_generator.py.
Lu & Wang · NeurIPS 2025 · 2025
Multi-agent KG enrichment that parallels SourcePrep’s multi-pass enrichment pipeline.
