H
Volume
AI & Retrieval
25 notes·14h to read
Other
24- Hybrid Search — BM25 + FAISS + RRFCombining keyword and vector retrieval — BM25 lexical scoring, FAISS dense-vector search, and merging both rankings with Reciprocal Rank Fusion.30m
- Structured Data Orchestration — Text to SQLWhy RAG can't do math over rows, the paradigm shift to a deterministic Text-to-SQL compiler, injecting schema (not data) into the prompt, locking the agent down with least privilege + a regex firewall, and self-correcting query loops.30m
- Multi-Source Retrieval — AlphaFund System DesignBuilding a finance agent that queries a SQL database and unstructured earnings transcripts in one sentence — a semantic router over tool-calling, self-correcting text-to-SQL, hybrid FAISS+BM25 RAG, plus the threat model and Phase 2 roadmap.30m
- Knowledge Graphs & Graph RAGWhy flat vector search misses relational reasoning — knowledge graph primitives (nodes, edges, properties), building a graph index with LLM extraction, and querying it with Graph RAG for multi-hop answers flat RAG can't reach.30m
- Knowledge Graph Primitives & Graph RAG with Neo4jHands-on construction of a property graph — nodes, edges, labels, and Cypher queries in Neo4j — then wiring it into a Graph RAG pipeline: entity extraction, community detection, and grounding LLM answers in structured graph context.30m
- Text → Graph — LLM Entity ExtractionBuilding the knowledge-graph engine (Neo4j, the Bolt driver, Docker, Cypher, APOC) and then automating its ingestion — turning thousands of messy documents into clean nodes and edges by using an LLM as a constrained reasoning parser: structured output, constrained decoding, ontology enforcement, and three-tier entity resolution.30m
- Graph RAG in Production → DSPyVectors stored inside Neo4j (HNSW + cosine), the three-phase hybrid retrieval (vector jump → 2-hop ego-graph traversal → LLM synthesis), taming super nodes with topological and semantic pruning, the hands-on Trinity Router, and the opening case for why hand-written prompts don't scale — enter DSPy.30m
- DSPy — Declarative OptimizationProgramming, not prompting. Treating prompts as tunable weights à la PyTorch: signatures as data contracts, modules the compiler can open, Predict vs. ChainOfThought, configuring the LM as "hardware," inspecting the compiled prompt, and the AlphaFund synthesizer that fuses SQL + FAISS + Neo4j into one executive summary.30m
- LangGraph — Agentic WorkflowsLangGraph fundamentals — building multi-step, stateful agent workflows on top of LangChain, defining nodes and edges, conditional branching, and orchestrating tool-calling agents as composable graphs.35m
- LangGraph, State Machines & Execution GraphsGoing deeper into LangGraph's state-machine model — defining typed state schemas, conditional edges as transition functions, cyclic graphs, checkpointing, human-in-the-loop breakpoints, and reasoning about agent control flow as a finite-state machine.35m
- State Machines & Execution GraphsWhy linear DAG pipelines break the moment an agent must decide — and how LangGraph replaces them with cyclic execution graphs built from three primitives: State, Nodes, and Edges, enabling loops, memory, and self-correction.35m
- The ReAct ArchitectureHow a language model stops guessing and starts working — interleaving reasoning with real tool calls in a Thought → Action → Observation loop until it has enough to answer, and why every production loop needs strict execution limits.35m
- Persistence & ThreadingHow a stateless agent learns to remember — saving a full state snapshot after every step, keyed to a conversation thread, so it survives restarts, reboots, and days-long gaps without leaking data across users.35m
- Context Management & SummarizationPersistence solves the no-memory problem; unbounded history then creates the too-much-memory problem. The fix is a context-management policy — verbatim, compress, discard — run in a dedicated maintenance node with custom reducers.35m
- Guardrails & Execution BoundingA production agent must stay bounded and stay skeptical — compressing old turns into a running summary instead of blindly trimming, and wrapping the model in guardrail nodes that mask PII and block prompt injections before it sees a single token.35m
- Tool Binding & ExecutionThe three-party contract between model, tool schema, and runtime — defining typed tools, binding only what an agent may use, executing requests in a tool node, filtering returned data, and turning failures into observations the loop can handle.35m
- The ReAct Agent — Production Loop SafetyA second pass over ReAct through the production failure modes — context bloat, duplicate tool calls, impure routers, unbounded cycles, escalation thresholds, and the circuit breakers that turn an elegant reasoning loop into a shippable agent.35m
- Safe Write Operations in Agentic SystemsWhy reads and writes need different architecture — write amplification, risk tiers, zero-trust credentials, idempotency keys, structured receipts, retry taxonomies, and defense in depth around a GitHub issue-creation tool.35m
- Secure Financial Write OperationsThe safe-write pattern under financial consequences — preventing duplicate spend, matching controls to irreversibility, keeping credentials outside the model, validating arguments, and returning traceable metadata instead of a void success response.35m
- Human in the Loop for Agentic Write OperationsWhere and how to pause an autonomous graph before consequences occur — blast-radius analysis, static versus dynamic breakpoints, interrupt-before semantics, checkpointed suspension, approve/edit/deny decisions, and reliable resumption.35m
- Persistent Agents & Context ManagementThe operational lifecycle of a long-running agent — checkpoint history, rewind and branching, interrupt/resume, persistent support threads, context growth, pruning versus summarization, and maintenance routing that keeps memory bounded.40m
- Securing Agentic AI SystemsA defense-in-depth reference for gullible agents — direct and indirect prompt injection, the relevant OWASP LLM risks, domain guardrails, classifier thresholds, PII detection and anonymization with Presidio, and the ingress-model-egress security sandwich.35m
- Multi-Agent TopologiesWhy the god-agent becomes an architectural liability — context dilution, quadratic tool confusion, three production topologies, formal handoffs, scoped state and tools, reducers for parallel writes, and subgraph boundaries that teams can test independently.40m
- The Supervisor OrchestratorA hub-and-spoke coordinator that reasons over full system state — deterministic versus semantic routing, structured decisions, pure routing functions, single-endpoint control, explicit FINISH semantics, delegation limits, and one auditable decision hub.35m