https://swapnanilsaha.com/blog/ Long-form engineering deep dives and research on retrieval, agent memory, and context engineering. Every number traces back to an instrumented run. en-us Mon, 20 Jul 2026 09:00:00 +0530 swapnanilsaha26@gmail.com (Swapnanil Saha) https://swapnanilsaha.com/blog/mcp-tool-adoption-agents/ https://swapnanilsaha.com/blog/mcp-tool-adoption-agents/ Mon, 20 Jul 2026 09:00:00 +0530 A controlled study of MCP tool adoption: 0 calls in 63 turns with 30 mentions of the tool — and what actually moves the number. https://swapnanilsaha.com/blog/deleting-ranking-heuristics-code-search/ https://swapnanilsaha.com/blog/deleting-ranking-heuristics-code-search/ Sun, 19 Jul 2026 09:00:00 +0530 3,817 lines of if/else, deleted in one commit — and the measured evidence for why query-keyword branches can never rank code search. https://swapnanilsaha.com/blog/embedding-dilution-code-search/ https://swapnanilsaha.com/blog/embedding-dilution-code-search/ Sat, 18 Jul 2026 09:00:00 +0530 The docstring paraphrased the query and it still ranked below 200 look-alikes. A measured post-mortem of embedding dilution — and the dual-vector fix it shipped. https://swapnanilsaha.com/blog/what-survives-compact-claude-code/ https://swapnanilsaha.com/blog/what-survives-compact-claude-code/ Fri, 17 Jul 2026 09:00:00 +0530 We forced 100+ compactions and graded, fact by fact, what the boundary keeps and what silently dies. The measured map, honest costs included. https://swapnanilsaha.com/blog/vectr-1-1-team-mode-seven-agent-test/ https://swapnanilsaha.com/blog/vectr-1-1-team-mode-seven-agent-test/ Sat, 11 Jul 2026 09:00:00 +0530 Team mode, auth, and encryption at rest — gated by a test where seven AI agents built and reviewed a product coordinating only through vectr's shared memory. https://swapnanilsaha.com/blog/vectr-v1-release-gate-honest-numbers/ https://swapnanilsaha.com/blog/vectr-v1-release-gate-honest-numbers/ Wed, 08 Jul 2026 09:00:00 +0530 A first-person dogfood gate caught two release-blocking bugs before the tag — and the honest cost numbers, wins and losses, that shipped with v1.0.0. https://swapnanilsaha.com/blog/four-families-llm-context-relief-eviction/ https://swapnanilsaha.com/blog/four-families-llm-context-relief-eviction/ Mon, 06 Jul 2026 09:00:00 +0530 Eviction, offload-and-recall, retrieval-over-stuffing, and subagent isolation — how to keep a coding agent from drowning in its own context, and why the four only work when composed. https://swapnanilsaha.com/blog/claude-code-hooks-deterministic-agent-memory/ https://swapnanilsaha.com/blog/claude-code-hooks-deterministic-agent-memory/ Sat, 04 Jul 2026 09:00:00 +0530 Events, the settings.json contract, exit codes — and a real production hook pipeline that injects working memory into an agent deterministically. https://swapnanilsaha.com/blog/building-vectr-part-3-benchmark-methodology-results/ https://swapnanilsaha.com/blog/building-vectr-part-3-benchmark-methodology-results/ Sun, 14 Jun 2026 09:00:00 +0530 How I benchmarked an AI code editor tool without fooling myself — methodology, the research vs implementation distinction, and what the CPython results actually show. https://swapnanilsaha.com/blog/building-vectr-part-2-working-memory-compact-survival/ https://swapnanilsaha.com/blog/building-vectr-part-2-working-memory-compact-survival/ Thu, 11 Jun 2026 09:00:00 +0530 Why /compact kills precision, what the KV cache actually is, and how I built a working memory layer that survives session boundaries — with the bugs that made it real. https://swapnanilsaha.com/blog/building-vectr-part-1-semantic-code-search/ https://swapnanilsaha.com/blog/building-vectr-part-1-semantic-code-search/ Tue, 09 Jun 2026 09:00:00 +0530 How I built a local semantic code indexer for AI editors — covering AST chunking, hybrid vector+BM25 search, symbol graphs, and why naive approaches break on real codebases. https://swapnanilsaha.com/blog/llm-context-window-token-budget-guide/ https://swapnanilsaha.com/blog/llm-context-window-token-budget-guide/ Tue, 26 May 2026 09:00:00 +0530 Token budget arithmetic for LLM context windows: why system prompts, tool schemas, RAG, and history drain your window — and four strategies to manage it. https://swapnanilsaha.com/blog/why-ai-code-assistants-waste-context-rag/ https://swapnanilsaha.com/blog/why-ai-code-assistants-waste-context-rag/ Sun, 24 May 2026 09:00:00 +0530 Context window stuffing silently degrades AI code assistant accuracy. Learn exactly why attention dilutes over long inputs and how RAG for code fixes it. https://swapnanilsaha.com/blog/text-embeddings-llms-rag-complete-guide/ https://swapnanilsaha.com/blog/text-embeddings-llms-rag-complete-guide/ Sat, 23 May 2026 09:00:00 +0530 From tokenization and transformers to cosine similarity, HNSW, RAG pipelines, and LLM training costs. With 7 interactive demos. https://swapnanilsaha.com/blog/dpdp-act-ai-llm-automation/ https://swapnanilsaha.com/blog/dpdp-act-ai-llm-automation/ Fri, 22 May 2026 09:00:00 +0530 What the DPDP Act actually requires, why shared inboxes fail, and how AI classifies requests, drafts multilingual replies, and generates audit-ready evidence — a deep technical guide.