CU-DBMI · d33bs · Jul 8, 2025 · Jul 10, 2025 · Jul 10, 2025
diff --git a/_posts/2025-07-08-llm_blog_post.md b/_posts/2025-07-08-llm_blog_post.md
@@ -0,0 +1,200 @@
+---
+title: "The Codex of Arcane Prompts: Questing Through the Realm of LLMs"
+author: dave-bunten
+tags:
+  - llms
+  - agents
+  - rag
+  - graphrag
+  - cot
+  - react
+---
+
+# The Codex of Arcane Prompts: Questing Through the Realm of LLMs
+
+{% include blog-post-intro.html %}
+
+## The Call to Adventure
+
+### Why LLMs Matter
+
+{% include figure.html image="images/word_wizards.png" width="80%"%}
+
+<!-- excerpt start -->
+In the kingdom of digital lore, LLMs are like legendary relics that breathe life into text.
+Forged through mighty Transformer architectures ([Vaswani et al. 2017](https://arxiv.org/abs/1706.03762), [Raffel et al. 2020](https://arxiv.org/abs/1910.10683)), they can summon summaries, conjure code, and translate runes from distant tongues—all with single incantations.
+As you learn to wield these arcane forces, fortresses of complexity crumble before you, and your quests become powered by computational "magic".
+
+Turn each page (or really, just scroll!) to discover core arcana, hidden libraries of lore, guild halls of platforms, the alchemist’s secret patterns, and a sorcerer’s arsenal of spells and runic glyphs.
+Paired with ancient scrolls (seminal papers), you’ll grasp not only how each enchantment works, but when to call it forth and why.
+<!-- excerpt end -->
+
+## The Tome of Foundations
+
+### The Essence of LLMs
+
+{% include figure.html image="images/llm_tapestry.png" width="80%"%}
+
+> **Runic Glyph: LLM**
+> A model that predicts language through self-attention and vast text training.
+
+Large language models (LLMs) leverage transformer (and other) models learn to predict the next token in the great tapestry of text, channeling self-attention to weave context.
+Whether wielding autoregressive GPT ([Brown et al. 2020](https://arxiv.org/abs/2005.14165)) or encoder–decoder T5 ([Raffel et al. 2020](https://arxiv.org/abs/1910.10683)), understanding their temperaments (generation vs. comprehension) ensures your projects hit their mark.
+
+### The Spellcraft of Prompting
+
+{% include figure.html image="images/wizard_speaking_to_tapestry.png" width="80%"%}
+
+> **Runic Glyph: Prompt**
+> Formatted text strings that guide LLMs magical generation.
+
+Prompts are your incantations: structured text containing instructions, contexts, and examples.
+From zero-shot to few-shot ([Brown et al. 2020](https://arxiv.org/abs/2005.14165)), these spells channel the LLM’s power.
+Craft them with clarity and brevity—too long, and tokens vanish into the abyss; too short, and your spell fizzles without effect.
+
+#### Prompting Rituals
+
+{% include figure.html image="images/llm_with_workflow.png" width="80%"%}
+
+##### Chain-of-Thought (CoT)
+
+{% include figure.html image="images/chain_of_thought.png" width="60%"%}
+
+> **Runic Glyph: Chain-of-Thought (CoT) **
+> A directive that unfolds the model’s intermediate reasoning steps.
+
+The CoT ___prompt-based reasoning method___ invites your LLM to “think aloud,” laying out each inferential step like glowing runes along a scroll ([Wei et al. 2022](https://arxiv.org/abs/2201.11903)).
+By prompting the model to enumerate its reasoning—“First I recall the ancient treaty… then I infer the binding clauses…”—you gain transparency into its logic and often a more accurate outcome.
+In practice, CoT spells are added as interim sections in your prompt, guiding the model through multi-step puzzles with clear, human-readable breadcrumbs.
+
+##### ReAct: Reason + Act
+
+{% include figure.html image="images/react_prompting.png" width="60%"%}
+
+> **Runic Glyph: ReAct**
+> A paradigm that interleaves model reasoning with external actions.
+
+ReAct weaves prompt-based reasoning runes with actionable rituals ([Yao et al. 2022](https://arxiv.org/abs/2210.03629)).
+As the LLM conjures its thought—“I should verify the current weather via the Weather API”—it pauses mid-incantation to call external familiars (APIs, functions, database queries), then resumes its reasoning with the newfound data.
+This blend of insight and interaction empowers dynamic workflows: imagine an agent that reasons, fetches a live stock price, then composes an investment recommendation—all within one continuous spell.
+
+Together, CoT and ReAct let you script complex magical rituals: you see *how* the spell unfolds and *when* it reaches beyond itself to summon additional data, creating both interpretability and power in your LLM applications.
+
+### Powering up your LLM as an agent
+
+{% include figure.html image="images/wizard_and_llm_with_familiar.png" width="80%"%}
+
+> **Runic Glyph: Agent**
+> A controller that orchestrates spells, keeps memory, and invokes external tools to enhance the use of LLMs.
+
+LLMs on their own have challenges with remembering context, sequencing their output, and executing external tools.
+Agents are your summoned familiars; magical constructs that combine reasoning and action for your LLMs.
+We can say that an LLM is "agentic" when it is enhanced with these further abilities that allow it to self-reason beyond what it is capable of alone.
+Agents can employ prompt-based reasoning methods as described above.
+
+#### Agent tools
+
+{% include figure.html image="images/react_with_tools.png" width="80%"%}
+
+> **Runic Glyph: Agent tools**
+> Tools provide your agent the ability to leverage additional capabilities such as executing a function or making an external request.
+
+No longer must you toil by candlelight, flipping through dusty parchments—these enchanted contrivances blend Thought and Deed into seamless ReAct rituals, summoning APIs, querying databases, and casting complex multi-step spells all in a breath.
+Verily, with Agent Tools at thy side, your LLM becomes not just a wise oracle, but a full-fledged magical order!
+
+#### MCP servers
+
+{% include figure.html image="images/react_with_mcp.png" width="60%"%}
+
+> **Runic Glyph: MCP Servers**
+> MCP servers offer a standardized way to remotely call services for agents.
+
+With MCP Servers at your command, you wield the power of a thousand wizards without so much as breaking a sweat—or a circuit.
+They are the unseen champions behind every majestic inference and data transformation, ensuring your digital sorcery proceeds swiftly and surely!
+
+## The Library of Lore
+
+{% include figure.html image="images/wizard_rag_tapestry_formation.png" width="80%"%}
+
+### Retrieval Data Vaults
+
+{% include figure.html image="images/rag_diagram.png" width="70%"%}
+
+> **Runic Glyph: RAG**
+> Retrieve → Condense → Generate: a pipeline for factual accuracy.
+
+RAG systems ([Lewis et al. 2020](https://arxiv.org/abs/2005.11401)) store document embeddings in vector vaults such as [Pinecone](https://www.pinecone.io/), [Weaviate](https://weaviate.io/), [Qdrant](https://qdrant.tech/), or embedded systems such as [DuckDB](https://duckdb.org/).
+At query time, you summon relevant scrolls (top‑k passages) and bind them into your spell, grounding the LLM in true lore and banishing hallucinations.
+
+### Knowledge-Crafted Graphs
+
+{% include figure.html image="images/graphrag_diagram.png" width="70%"%}
+
+> **Runic Glyph: GraphRAG**
+> Embedding retrieval enchanted with graph-based evidence chains.
+
+GraphRAG ([Chen et al. 2021](https://arxiv.org/abs/2109.01117)) blends vector summoning with graph traversals, exploring relationships among entities like a scholar poring over ancient tomes.
+This multi-hop reasoning provides provenance and deeper insights, ideal for complex quests in biomedical or enterprise realms.
+
+## The Guildhall of Platforms
+
+{% include figure.html image="images/external_vs_local_llm.png" width="80%"%}
+
+### Hosted APIs
+
+{% include figure.html image="images/cloud_llms.png" width="40%"%}
+
+> **Runic Glyph: Hosted APIs**
+> A managed endpoint for commanding pre-trained LLMs.
+
+Closed-source oracles—OpenAI, Anthropic, Google Vertex AI—offer turnkey summoning circles.
+They manage scaling, compliance, and SLAs, but heed their pricing scrolls and data policies before pledging your allegiance.
+
+### Self-hosted APIs
+
+{% include figure.html image="images/castle_llm.png" width="40%"%}
+
+> **Runic Glyph: Self-hosted APIs**
+> Self-host your LLM relics for ultimate sovereignty.
+
+Open-source forges such as [Ollama](https://github.com/ollama/ollama), [llama.cpp](https://github.com/ggerganov/llama.cpp), [vLLM](https://github.com/vllm-project/vllm) let you craft and run models on your own hearth.
+While you shoulder hardware and ops burdens, you gain full control, cost predictability, and on-prem solace.
+
+## The Alchemist’s Workshop
+
+{% include figure.html image="images/alchemist_experiment.png" width="80%"%}
+
+### LLM Orchestrators
+
+{% include figure.html image="images/agent_chaining.png" width="60%"%}
+
+{% include figure.html image="images/agent_coordinator.png" width="60%"%}
+
+> **Runic Glyph: LLM Orchestrator**
+> A framework for chaining prompts and architecting agents.
+
+Invoke [LangGraph](https://github.com/langchain-ai/langgraph), [LlamaIndex](https://github.com/run-llama/llama_index), [Haystack](https://github.com/deepset-ai/haystack), or the [Google AI Developer Kit](https://github.com/google/adk-docs) (google-adk) to access pre-made chains, retrievers, and LLM adapters.
+These libraries are like scrolls of ancient knowledge, saving you countless hours of boilerplate incantations.
+
+### Interactive web interface packages
+
+{% include figure.html image="images/web_interface_packages.png" width="70%"%}
+
+> **Runic Glyph: Web interface packages**
+> Packages which allow you to quickly build and share web interfaces, including chat-based interactions with agents.
+
+Whether through conversational UIs ([Streamlit](https://github.com/streamlit/streamlit), [Gradio](https://github.com/gradio-app/gradio)) or direct API rituals (REST, SDKs), choose the conduit that best fits your ritual circle.
+Many begin with a demo UI, then migrate enchanted logic into backend services for production-grade robustness.
+
+## The Final Ritual
+
+### Wizardry Recap
+You’ve mastered foundational spells, explored hidden libraries of lore, navigated guild halls of platforms, wielded alchemical patterns, and equipped a sorcerer’s arsenal.
+You know the runic glyphs and have bound them to seminal scrolls.
+
+### Further Scrolls to Study
+Continue your journey by reading the original Transformer codex ([Vaswani et al. 2017](https://arxiv.org/abs/1706.03762)), RAG rituals ([Lewis et al. 2020](https://arxiv.org/abs/2005.11401)), the Chain-of-Thought codex ([Wei et al. 2022](https://arxiv.org/abs/2201.11903)), and ReAct incantations ([Yao et al. 2022](https://arxiv.org/abs/2210.03629)).
+
+### Your Grand Quest
+Forge your own grimoire: spin up a RAG service or an agent familiar, document your arcane experiments, and share your triumphs in a blog chronicle or lightning talk.
+The realm of LLM sorcery awaits your legend!
diff --git a/images/agent_chaining.png b/images/agent_chaining.png
diff --git a/images/agent_coordinator.png b/images/agent_coordinator.png
diff --git a/images/alchemist_experiment.png b/images/alchemist_experiment.png
diff --git a/images/castle_llm.png b/images/castle_llm.png
diff --git a/images/chain_of_thought.png b/images/chain_of_thought.png
diff --git a/images/cloud_llms.png b/images/cloud_llms.png
diff --git a/images/external_vs_local_llm.png b/images/external_vs_local_llm.png
diff --git a/images/graphrag_diagram.png b/images/graphrag_diagram.png
diff --git a/images/llm_tapestry.png b/images/llm_tapestry.png
diff --git a/images/llm_with_workflow.png b/images/llm_with_workflow.png
diff --git a/images/rag_diagram.png b/images/rag_diagram.png
diff --git a/images/react_prompting.png b/images/react_prompting.png
diff --git a/images/react_with_mcp.png b/images/react_with_mcp.png
diff --git a/images/react_with_tools.png b/images/react_with_tools.png
diff --git a/images/web_interface_packages.png b/images/web_interface_packages.png
diff --git a/images/wizard_and_llm_with_familiar.png b/images/wizard_and_llm_with_familiar.png
diff --git a/images/wizard_rag_tapestry_formation.png b/images/wizard_rag_tapestry_formation.png
diff --git a/images/wizard_speaking_to_tapestry.png b/images/wizard_speaking_to_tapestry.png
diff --git a/images/word_wizards.png b/images/word_wizards.png