/** * PilotMaster Comprehensive Knowledge & Terminology System * Strictly scoped by Page and Mode with intelligent multi-token fuzzy search. */ export const GLOSSARY_CATEGORIES = [ "All Categories", "Foundation Models", "Embeddings & Vectors", "Rerankers", "Retrieval Strategies", "Chunking Strategies", "Query Enhancements", "Observability & Telemetry", "Benchmarking & Evaluation", "Platform & Architecture", ]; export const GLOSSARY_TERMS = [ // ───────────────────────────────────────────────────────────────────────────── // 1. FOUNDATION MODELS (LLMs) // ───────────────────────────────────────────────────────────────────────────── { id: "model_gpt_oss_120b", title: "GPT-OSS 120B", aliases: ["gpt-oss-120b", "openai/gpt-oss-120b", "gpt oss 120b", "frontier model", "120b", "large model"], pages: ["docpilot", "gaugepilot", "home"], modes: ["all"], category: "Foundation Models", definition: "A high-intelligence 120-billion parameter frontier open-weights LLM running with sub-second token streaming via Groq LPU inference.", whyItMatters: "Serves as the flagship synthesis engine for complex document reasoning, multi-document synthesis, structured GFM markdown tables, and comprehensive answer grounding.", location: "DocPilot Model selector dropdown, GaugePilot model matrix, and Home system badge.", }, { id: "model_gpt_oss_20b", title: "GPT-OSS 20B", aliases: ["gpt-oss-20b", "openai/gpt-oss-20b", "gpt oss 20b", "fast model", "20b", "efficient rag"], pages: ["docpilot", "gaugepilot"], modes: ["all"], category: "Foundation Models", definition: "An ultra-fast, compact 20-billion parameter neural model optimized for high-throughput, low-latency question answering and preliminary query transformations.", whyItMatters: "Delivers 3-5x faster time-to-first-token (TTFT) with low token cost while maintaining strong factual adherence.", location: "DocPilot Model selector and GaugePilot Experiment Setup.", }, { id: "model_qwen_3_6_27b", title: "Qwen 3.6 27B", aliases: ["qwen/qwen3.6-27b", "qwen 3.6", "qwen", "qwen27b", "alibaba", "qwen3.6"], pages: ["docpilot", "gaugepilot"], modes: ["all"], category: "Foundation Models", definition: "Alibaba's advanced 27-billion parameter multilingual foundation model with specialized capabilities in code understanding, mathematical reasoning, and complex table analysis.", whyItMatters: "Excels at parsing technical manuals, structured CSVs, financial tables, and multilingual knowledge bases.", location: "DocPilot Model selector and GaugePilot model grid.", }, { id: "model_deepseek_r1_70b", title: "DeepSeek R1 70B (Distill Llama)", aliases: ["deepseek-r1-distill-llama-70b", "deepseek r1", "deepseek", "r1", "reasoning model", "cot", "chain of thought"], pages: ["docpilot", "gaugepilot"], modes: ["all"], category: "Foundation Models", definition: "DeepSeek's distilled 70-billion parameter reasoning model with reinforcement-learned deep chain-of-thought (CoT) problem solving capabilities.", whyItMatters: "Performs deep internal deduction to reconcile contradictory passages, resolve multi-hop logic questions, and verify mathematical computations before synthesizing answers.", location: "DocPilot Model selector and GaugePilot benchmark matrix.", }, { id: "model_llama_3_3_70b", title: "Meta Llama 3.3 70B", aliases: ["llama-3.3-70b-versatile", "llama 3.3 70b", "llama", "meta llama", "llama3"], pages: ["gaugepilot", "docpilot"], modes: ["all"], category: "Foundation Models", definition: "Meta's flagship open-weights instruction-tuned model with 128k context window support and high general benchmark performance.", whyItMatters: "Industry standard benchmark reference for instruction following, faithfulness, and high-quality factual RAG answers.", location: "GaugePilot model comparison and benchmark candidate list.", }, { id: "model_mixtral_8x7b", title: "Mixtral 8x7B (MoE)", aliases: ["mixtral-8x7b-32768", "mixtral", "mistral", "moe", "mixture of experts"], pages: ["gaugepilot"], modes: ["exp"], category: "Foundation Models", definition: "A sparse Mixture-of-Experts (MoE) architecture activating 13B parameters per token out of 47B total parameters with a 32k context window.", whyItMatters: "Provides top-tier inference speed with frontier-level quality across broad European languages and code syntax.", location: "GaugePilot model options and comparative leaderboards.", }, { id: "model_gemini_1_5", title: "Google Gemini 1.5 Series", aliases: ["gemini", "gemini-1.5-flash", "gemini-1.5-pro", "google gemini"], pages: ["gaugepilot"], modes: ["exp"], category: "Foundation Models", definition: "Google's multimodal transformer models featuring extended context windows (up to 1M+ tokens) and rapid cross-attention retrieval.", whyItMatters: "Serves as an external benchmark target for long-context recall and multi-modal document reasoning.", location: "GaugePilot evaluation leaderboards and radar profile.", }, // ───────────────────────────────────────────────────────────────────────────── // 2. EMBEDDINGS & VECTORS // ───────────────────────────────────────────────────────────────────────────── { id: "emb_mpnet_base_v2", title: "all-mpnet-base-v2 (768 Dim)", aliases: ["all-mpnet-base-v2", "mpnet", "mpnet base", "sentence transformers", "768"], pages: ["docpilot", "gaugepilot", "home"], modes: ["all"], category: "Embeddings & Vectors", definition: "A 768-dimensional sentence embedding model pre-trained on over 1 billion sentence pairs using masked and permuted language modeling.", whyItMatters: "The default balanced general-purpose embedding model for PilotMaster, offering high semantic clustering accuracy across varied domains.", location: "DocPilot Embedding dropdown and PilotMaster Home footer.", }, { id: "emb_minilm_l6_v2", title: "all-MiniLM-L6-v2 (384 Dim)", aliases: ["all-MiniLM-L6-v2", "minilm-l6", "minilm 384", "fast embeddings", "384"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Embeddings & Vectors", definition: "A compact 384-dimensional embedding model optimized for lightning-fast encoding and low memory footprint with 5x throughput over standard models.", whyItMatters: "Ideal for edge deployments, rapid indexing of large document batches, and low-latency search systems.", location: "DocPilot Embedding selector in Experimental Mode.", }, { id: "emb_minilm_l12_v2", title: "all-MiniLM-L12-v2 (384 Dim)", aliases: ["all-MiniLM-L12-v2", "minilm-l12", "minilm 12"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Embeddings & Vectors", definition: "A 12-layer variant of MiniLM offering higher semantic resolution than L6 while retaining a compact 384-dimensional vector size.", whyItMatters: "Provides an optimal sweet spot between encoding speed and retrieval precision.", location: "DocPilot Embedding selector.", }, { id: "emb_bge_large_en", title: "BAAI BGE-Large (bge-large-en-v1.5 · 1024 Dim)", aliases: ["bge-large-en-v1.5", "bge-large", "bge", "baai", "1024"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Embeddings & Vectors", definition: "A state-of-the-art 1024-dimensional dense text embedding model developed by the Beijing Academy of Artificial Intelligence (BAAI).", whyItMatters: "Ranks near the top of the Massive Text Embedding Benchmark (MTEB) for complex semantic search and question-to-passage matching.", location: "DocPilot Embedding selector dropdown in Experimental Mode.", }, { id: "emb_gte_large", title: "Thenlper GTE-Large (1024 Dim)", aliases: ["gte-large", "gte", "thenlper", "general text embeddings"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Embeddings & Vectors", definition: "A 1024-dimensional general text embedding model trained on large-scale multi-domain web corpora for multi-task semantic ranking.", whyItMatters: "High resilience across noisy documents, technical manuals, and conversational dialogues.", location: "DocPilot Embedding dropdown.", }, { id: "emb_openai_small", title: "OpenAI text-embedding-3-small (1536 Dim)", aliases: ["text-embedding-3-small", "openai embedding", "1536", "embedding 3 small"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Embeddings & Vectors", definition: "OpenAI's efficient 1536-dimensional dense embedding model with native dimensional reduction support.", whyItMatters: "Standard cloud-hosted embedding baseline for commercial enterprise RAG architectures.", location: "DocPilot Embedding dropdown.", }, { id: "emb_openai_large", title: "OpenAI text-embedding-3-large (3072 Dim)", aliases: ["text-embedding-3-large", "embedding 3 large", "3072", "high dimensional embedding"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Embeddings & Vectors", definition: "OpenAI's flagship 3072-dimensional embedding model delivering maximum semantic nuance and fine-grained classification accuracy.", whyItMatters: "Distinguishes subtle distinctions in legal, medical, and specialized technical literature.", location: "DocPilot Embedding dropdown in Experimental Mode.", }, // ───────────────────────────────────────────────────────────────────────────── // 3. RERANKERS // ───────────────────────────────────────────────────────────────────────────── { id: "reranker_none", title: "Identity Reranker (None)", aliases: ["none", "identity reranker", "no reranker", "raw retrieval"], pages: ["docpilot", "gaugepilot"], modes: ["all"], category: "Rerankers", definition: "Bypasses the second-stage neural reranking pass, passing candidate chunks directly in their raw vector/lexical retrieval order to the LLM prompt.", whyItMatters: "Minimizes pipeline latency when speed is paramount and first-stage vector retrieval confidence is already high.", location: "DocPilot Reranker dropdown -> 'None'.", }, { id: "reranker_minilm", title: "MiniLM Cross-Encoder (ms-marco-MiniLM-L-6-v2)", aliases: ["minilm", "cross-encoder", "cross encoder", "ms-marco", "minilm cross encoder"], pages: ["docpilot", "tracepilot", "gaugepilot"], modes: ["exp", "all"], category: "Rerankers", definition: "A neural cross-encoder that jointly encodes query and passage tokens across all transformer self-attention layers, computing a direct relevance logit score.", whyItMatters: "Dramatically outperforms bi-encoder similarity search by capturing deep token-to-token semantic interactions, eliminating false-positive chunk matches.", location: "DocPilot Reranker selector and TracePilot chunk diagnostics.", }, { id: "reranker_tinybert", title: "TinyBERT Reranker", aliases: ["tinybert", "tiny bert", "lightweight reranker", "fast reranker"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Rerankers", definition: "A distilled 4-layer transformer reranker providing neural ranking capabilities with sub-10 millisecond inference overhead.", whyItMatters: "Enables neural reranking on resource-constrained or latency-critical production paths.", location: "DocPilot Reranker dropdown in Experimental Mode.", }, { id: "reranker_bge_large", title: "BGE Large Reranker (bge-reranker-large)", aliases: ["bge-large", "bge reranker", "baai reranker", "high precision reranker"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Rerankers", definition: "A heavyweight 560M-parameter cross-encoder trained explicitly on question-passage relevance benchmarks.", whyItMatters: "Achieves state-of-the-art accuracy in ranking the single most relevant chunk to rank #1.", location: "DocPilot Reranker dropdown.", }, { id: "reranker_bge_m3", title: "BGE M3 Multilingual Reranker", aliases: ["bge-m3", "bge m3", "multilingual reranker"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Rerankers", definition: "A cross-lingual neural reranker trained on over 100 languages with support for long input sequences up to 8192 tokens.", whyItMatters: "Essential for cross-lingual enterprise search where queries in one language match documents in another.", location: "DocPilot Reranker dropdown.", }, { id: "reranker_flashrank", title: "FlashRank Quantized Reranker", pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Rerankers", definition: "An ultra-fast, CPU-optimized, quantized neural reranker designed for sub-millisecond reranking without GPU dependencies.", whyItMatters: "Delivers 85-90% of full cross-encoder accuracy with 5-10x speedup and minimal RAM usage.", location: "DocPilot Reranker dropdown.", }, { id: "reranker_cohere", title: "Cohere Rerank API", aliases: ["cohere", "cohere rerank", "cohere api"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Rerankers", definition: "An enterprise cloud-hosted reranking endpoint optimized for business documents, tables, and long-form context reordering.", whyItMatters: "Provides top-tier commercial reranking accuracy across varied formats.", location: "DocPilot Reranker dropdown.", }, // ───────────────────────────────────────────────────────────────────────────── // 4. RETRIEVAL STRATEGIES // ───────────────────────────────────────────────────────────────────────────── { id: "strat_vector_search", title: "Dense Vector Search (FAISS)", aliases: ["dense", "vector search", "dense vector", "faiss", "semantic search", "cosine similarity"], pages: ["docpilot", "tracepilot", "gaugepilot"], modes: ["all"], category: "Retrieval Strategies", definition: "Mathematical search representing documents and queries as high-dimensional vectors and finding nearest neighbors via inner-product or cosine similarity.", whyItMatters: "Understands meaning, context, synonyms, and intent even when the user query uses different words than the document.", formula: "cosine_sim(q, d) = (q · d) / (||q|| · ||d||)", location: "DocPilot Retrieval selector and TracePilot span waterfall.", }, { id: "strat_bm25", title: "BM25 Keyword Search", aliases: ["bm25", "lexical", "keyword search", "sparse retrieval", "inverted index"], pages: ["docpilot", "tracepilot", "gaugepilot"], modes: ["exp", "all"], category: "Retrieval Strategies", definition: "A probabilistic lexical ranking function that scores passages based on term frequency (TF), inverse document frequency (IDF), and document length normalization.", whyItMatters: "Essential for finding exact product names, error codes, part numbers, and verbatim code signatures where dense embeddings may fail.", location: "DocPilot Retrieval selector and TracePilot Lexical concordance card.", }, { id: "strat_hybrid", title: "Hybrid Search (Dense Vector + BM25 + RRF)", aliases: ["hybrid", "hybrid search", "ensemble retrieval", "dense + bm25", "rrf"], pages: ["docpilot", "tracepilot", "gaugepilot"], modes: ["exp", "all"], category: "Retrieval Strategies", definition: "Simultaneously executes dense semantic vector search and sparse BM25 keyword search, merging the two candidate ranked lists using Reciprocal Rank Fusion (RRF).", whyItMatters: "The gold standard for RAG retrieval: combines deep semantic intent matching with precision keyword lookups for maximum recall and accuracy.", formula: "RRF(d) = Σ [ 1 / (k + rank_i(d)) ] where k=60", location: "DocPilot Retrieval selector in Experimental Mode.", }, { id: "strat_contextual_compression", title: "Contextual Compression", aliases: ["compression", "context compression", "passage extractor", "sentence selector"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Retrieval Strategies", definition: "Extracts only the query-relevant sentences from retrieved passages, discarding irrelevant surrounding text before prompt injection.", whyItMatters: "Reduces prompt token costs, minimizes context window bloat, and mitigates the 'lost in the middle' attention degradation in LLMs.", location: "DocPilot Experimental pipeline configuration.", }, // ───────────────────────────────────────────────────────────────────────────── // 5. CHUNKING STRATEGIES // ───────────────────────────────────────────────────────────────────────────── { id: "chunk_parent_child", title: "Parent-Child Chunking (1200 / 300)", aliases: ["parent-child", "parent child", "hierarchical chunking", "1200 / 300"], pages: ["docpilot", "gaugepilot"], modes: ["all"], category: "Chunking Strategies", definition: "Splits documents into small 'child' chunks (300 chars) for high-precision vector search, but resolves to their larger 'parent' chunk (1200 chars) when constructing the LLM generation prompt.", whyItMatters: "Solves the fundamental RAG trade-off: small chunks maximize vector retrieval precision; large parent chunks give the LLM full context without truncation.", location: "DocPilot Chunking selector dropdown and ingestion pipeline.", }, { id: "chunk_recursive", title: "Recursive Character Splitting", aliases: ["recursive", "recursive character", "paragraph splitting"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Chunking Strategies", definition: "Splits documents hierarchically using ordered delimiters (double newlines, single newlines, spaces, characters) to maintain paragraph and sentence integrity.", whyItMatters: "Standard robust text chunking strategy that prevents split sentences and preserves formatting.", location: "DocPilot Chunking dropdown in Experimental Mode.", }, { id: "chunk_fixed", title: "Fixed Window Chunking (500c / 50c)", aliases: ["fixed", "fixed window", "character window", "500c"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Chunking Strategies", definition: "Splits text into fixed character counts (e.g. 500 characters) with a sliding overlap window (50 characters).", whyItMatters: "Deterministic and fast, ensuring uniform vector chunk distribution across large text corpora.", location: "DocPilot Chunking dropdown.", }, { id: "chunk_token", title: "Token-Based Chunking (256t)", aliases: ["token", "token-based", "token chunking", "256t", "bpe"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Chunking Strategies", definition: "Splits text strictly along Byte-Pair Encoding (BPE) tokenizer boundaries rather than raw character counts.", whyItMatters: "Guarantees exact context window budget management and prevents mid-subword token corruption.", location: "DocPilot Chunking dropdown.", }, { id: "chunk_semantic", title: "Semantic Similarity Chunking", aliases: ["semantic chunking", "cosine chunking", "topic shift chunking"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Chunking Strategies", definition: "Computes vector embeddings for consecutive sentences and splits the text dynamically at statistical drops in cosine similarity (topic shifts).", whyItMatters: "Ensures each chunk represents a single coherent concept or topic.", location: "DocPilot Chunking dropdown in Experimental Mode.", }, { id: "chunk_contextual", title: "Contextual Chunking (Contextual Retrieval)", aliases: ["contextual chunking", "contextual retrieval", "situating context", "chunk prefix", "anthropic contextual"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Chunking Strategies", definition: "Leverages a fast LLM during document ingestion to generate a concise 20–40 word situating context prefix prepended to each chunk (e.g. '[Context: From SEC 10-K report of Acme Corp 2024...]') before computing vector embeddings and BM25 indexes.", whyItMatters: "Eliminates the 'isolated chunk problem' where extracted passages mention facts or numbers without identifying the overarching company, timeframe, or document topic.", location: "DocPilot Chunking selector -> 'Contextual Chunking' and GaugePilot Experiment Setup.", }, { id: "chunk_structure_aware", title: "Structure-Aware Chunking", aliases: ["structure-aware", "structure aware chunking", "markdown chunking", "html chunking", "heading hierarchy", "ast chunking"], pages: ["docpilot", "gaugepilot"], modes: ["exp", "all"], category: "Chunking Strategies", definition: "Parses document syntax (Markdown #, ##, ### headers, HTML

-

elements, code fences, and markdown tables) into logical semantic sections, attaching breadcrumb header paths (e.g. '[Architecture > Database Layer > FAISS]') to sub-chunks.", whyItMatters: "Guarantees that tables, code snippets, and nested section headings are preserved intact without being fractured across arbitrary character or token cutoffs.", location: "DocPilot Chunking selector -> 'Structure-Aware' and GaugePilot Experiment Setup.", }, // ───────────────────────────────────────────────────────────────────────────── // 6. 11 QUERY ENHANCEMENTS (LAB / EXP) // ───────────────────────────────────────────────────────────────────────────── { id: "enh_condensation", title: "Query Condensation / Contextualization", aliases: ["query condensation", "conversational contextualization", "standalone query", "followup rewrite"], pages: ["docpilot", "tracepilot", "gaugepilot"], modes: ["exp", "all"], category: "Query Enhancements", definition: "Transforms follow-up conversational questions ('What about its pricing?') into fully self-contained standalone search queries ('What is the pricing model of PilotMaster?').", whyItMatters: "Prevents retrieval failure on follow-up questions where vector search would otherwise lack context from prior conversation turns.", location: "DocPilot Enhancements dropdown -> Context Preparation.", }, { id: "enh_coreference", title: "Coreference Resolution", aliases: ["coreference", "coreference resolution", "pronoun resolution", "entity resolver"], pages: ["docpilot", "tracepilot", "gaugepilot"], modes: ["exp", "all"], category: "Query Enhancements", definition: "Replaces ambiguous pronouns (it, they, that company, former/latter) with explicit named entities extracted from conversation history.", whyItMatters: "Guarantees vector searches target specific entity names rather than vague pronoun embeddings.", location: "DocPilot Enhancements dropdown -> Coreference Resolution.", }, { id: "enh_rewrite", title: "Query Rewrite", aliases: ["query rewrite", "query refactor", "clarification"], pages: ["docpilot", "tracepilot", "gaugepilot"], modes: ["exp", "all"], category: "Query Enhancements", definition: "Rewrites verbose, informal, or grammatically noisy user prompts into clean, concise, retrieval-optimized search statements.", whyItMatters: "Improves retrieval recall by removing conversational filler words that dilute vector similarity.", location: "DocPilot Enhancements dropdown -> Query Rewrite.", }, { id: "enh_subquery", title: "Sub-Query Generation", aliases: ["subquery", "sub-query generation", "query decomposition", "multi-part query"], pages: ["docpilot", "tracepilot", "gaugepilot"], modes: ["exp", "all"], category: "Query Enhancements", definition: "Decomposes complex multi-part questions (e.g. 'Compare the battery, price, and warranty of Model A vs B') into independent parallel sub-queries.", whyItMatters: "Ensures complete recall across all sub-questions rather than retrieving a single compromise vector.", location: "DocPilot Enhancements dropdown -> Structuring & Routing.", }, { id: "enh_metadata", title: "Metadata Filter Extraction", aliases: ["metadata filter", "metadata filter extraction", "structured filters", "date extraction"], pages: ["docpilot", "tracepilot", "gaugepilot"], modes: ["exp", "all"], category: "Query Enhancements", definition: "Extracts structured filters (dates, authors, departments, file types) from natural language to apply hard constraints to vector queries.", whyItMatters: "Narrows the search space instantly, filtering out irrelevant timeframes and document types before similarity calculation.", location: "DocPilot Enhancements dropdown -> Metadata Filter Extraction.", }, { id: "enh_routing", title: "Query Routing", aliases: ["query routing", "index routing", "intent classification"], pages: ["docpilot", "tracepilot", "gaugepilot"], modes: ["exp", "all"], category: "Query Enhancements", definition: "Classifies user intent and routes queries to specialized vector indices, tables, or domain pipelines.", whyItMatters: "Directs code queries to code indices and financial queries to numerical tables.", location: "DocPilot Enhancements dropdown -> Query Routing.", }, { id: "enh_hyde", title: "HyDE (Hypothetical Document Embeddings)", aliases: ["hyde", "hypothetical document", "synthetic answer embedding"], pages: ["docpilot", "tracepilot", "gaugepilot"], modes: ["exp", "all"], category: "Query Enhancements", definition: "Uses an LLM to generate a hypothetical ideal answer to the question, then performs vector search using the embedding of that synthetic document.", whyItMatters: "Bridges the semantic gap between questions and answers: questions look very different from answers in vector space, but a hypothetical answer looks almost identical to true answer chunks.", location: "DocPilot Enhancements dropdown -> HyDE.", }, { id: "enh_multiquery", title: "Multi-Query Expansion", aliases: ["multi-query", "multi query expansion", "query variations", "parallel search"], pages: ["docpilot", "tracepilot", "gaugepilot"], modes: ["exp", "all"], category: "Query Enhancements", definition: "Generates 3 to 5 semantically diverse rephrasings of the user query from different perspectives and executes concurrent retrieval across all variations.", whyItMatters: "Overcomes the sensitivity of vector similarity to specific word choices, ensuring maximum candidate recall.", location: "DocPilot Enhancements dropdown -> Multi-Query Expansion.", }, { id: "enh_ragfusion", title: "RAG-Fusion", aliases: ["rag-fusion", "rag fusion", "rrf multi-query", "fusion retrieval"], pages: ["docpilot", "tracepilot", "gaugepilot"], modes: ["exp", "all"], category: "Query Enhancements", definition: "Generates multiple query formulations, retrieves candidate documents for each, and applies Reciprocal Rank Fusion (RRF) to merge and re-rank the union of results.", whyItMatters: "Combines the recall benefits of multi-query expansion with the ranking stability of rank-aggregation algorithms.", location: "DocPilot Enhancements dropdown -> RAG-Fusion.", }, { id: "enh_stepback", title: "Step-Back Prompting", aliases: ["step-back", "step back prompting", "conceptual abstraction", "high level query"], pages: ["docpilot", "tracepilot", "gaugepilot"], modes: ["exp", "all"], category: "Query Enhancements", definition: "Generates a broader, higher-level conceptual query (stepping back from specific details) to retrieve foundational domain principles.", whyItMatters: "Equips the LLM with the necessary background theories and rules needed to reason through complex questions.", location: "DocPilot Enhancements dropdown -> Step-Back Prompting.", }, { id: "enh_keyword_exp", title: "Query Keyword Expansion", aliases: ["keyword expansion", "synonym injection", "domain terminology expansion"], pages: ["docpilot", "tracepilot", "gaugepilot"], modes: ["exp", "all"], category: "Query Enhancements", definition: "Enriches queries with domain-specific terminology, technical acronyms, and lexical variants prior to BM25 search.", whyItMatters: "Dramatically increases keyword hit rates without requiring users to know exact technical nomenclature.", location: "DocPilot Enhancements dropdown -> Query Keyword Expansion.", }, // ───────────────────────────────────────────────────────────────────────────── // 7. OBSERVABILITY & TELEMETRY (TRACEPILOT) // ───────────────────────────────────────────────────────────────────────────── { id: "obs_trace_id", title: "Trace ID & Trace Lifecycle", aliases: ["trace id", "trace", "uuid", "execution lifecycle", "trace lifecycle"], pages: ["tracepilot"], modes: ["all"], category: "Observability & Telemetry", definition: "A globally unique identifier (UUID) assigned to each query execution, connecting all preprocessing, retrieval, reranking, and generation telemetry.", whyItMatters: "Provides end-to-end auditability and debugging capabilities for any individual query across the platform.", location: "TracePilot sidebar item header and main telemetry header.", }, { id: "obs_spans", title: "Execution Spans (Root & Child)", aliases: ["spans", "root span", "child spans", "span hierarchy", "timeline waterfall"], pages: ["tracepilot"], modes: ["all"], category: "Observability & Telemetry", definition: "Hierarchical execution units measuring precise timestamps, durations, inputs, and outputs for every pipeline sub-stage (Preprocessing, Retrieval, Rerank, Prompt, Generation).", whyItMatters: "Identifies exact latency bottlenecks, such as slow embedding calls or oversized prompt payloads.", location: "TracePilot Timeline & Spans view.", }, { id: "obs_dag", title: "Execution DAG / Graph View", aliases: ["dag", "execution dag", "graph view", "pipeline graph", "visual execution"], pages: ["tracepilot"], modes: ["exp"], category: "Observability & Telemetry", definition: "A visual Directed Acyclic Graph (DAG) charting the live flow of data between query enhancement branches, parallel retrieval nodes, fusion points, and synthesis stages.", whyItMatters: "Makes complex branching multi-query and hybrid retrieval pipelines intuitive to understand and debug.", location: "TracePilot Experimental Mode 'DAG Graph' tab.", }, { id: "obs_replay", title: "Replay Execution Engine", aliases: ["replay", "replay trace", "re-execute", "determinism test"], pages: ["tracepilot"], modes: ["all"], category: "Observability & Telemetry", definition: "Re-executes a historical trace with identical parameters, query, and documents, measuring performance variance and verifying determinism.", whyItMatters: "Crucial for reproducing bugs, validating pipeline improvements, and measuring model consistency.", location: "TracePilot 'Replay Trace' button in the trace details view.", }, { id: "obs_tokens_cost", title: "Token Usage & Cost Telemetry", aliases: ["tokens", "token usage", "prompt tokens", "completion tokens", "cost", "usd cost"], pages: ["tracepilot", "docpilot"], modes: ["all"], category: "Observability & Telemetry", definition: "Precise accounting of Prompt Tokens (input context), Completion Tokens (generated output), Total Tokens, and estimated USD cost based on model pricing.", whyItMatters: "Enables production cost budgeting, token optimization, and context window efficiency monitoring.", location: "TracePilot metrics bar and DocPilot response metadata.", }, { id: "obs_consensus", title: "Consensus & Agreement Score", aliases: ["consensus", "agreement score", "retrieval consensus", "semantic overlap"], pages: ["tracepilot"], modes: ["exp"], category: "Observability & Telemetry", definition: "Measures the degree of semantic and chunk overlap across multiple retrieval branches (e.g. between Dense and Lexical search, or across Multi-Query formulations).", whyItMatters: "High consensus indicates strong retrieval confidence, while low consensus warns of ambiguity in the source corpus.", location: "TracePilot Experimental Overview and Telemetry Cards.", }, { id: "obs_concordance", title: "Lexical vs Semantic Concordance", aliases: ["concordance", "lexical vs semantic", "bm25 vs dense concordance"], pages: ["tracepilot"], modes: ["exp"], category: "Observability & Telemetry", definition: "A comparative metric showing whether BM25 keyword search and Dense Vector embeddings agreed on the same top candidate chunks.", whyItMatters: "Diagnoses whether a query was primarily resolved by keyword matches or conceptual semantic similarity.", location: "TracePilot Concordance analysis section.", }, { id: "obs_risk_index", title: "Hallucination & Risk Index", aliases: ["risk index", "hallucination risk", "risk score", "guardrail"], pages: ["tracepilot"], modes: ["exp"], category: "Observability & Telemetry", definition: "An automated risk score rating the likelihood that generated response claims deviate from the provided source context chunks.", whyItMatters: "Acts as an automated guardrail flagging high-risk responses before they reach end users in mission-critical applications.", location: "TracePilot risk banner and overview telemetry.", }, { id: "obs_diagnostics", title: "Retrieval Diagnostics & Chunk Lineage", aliases: ["lineage", "chunk diagnostics", "dense score", "bm25 score", "rrf score", "reranker score"], pages: ["tracepilot"], modes: ["all"], category: "Observability & Telemetry", definition: "A multi-stage score progression tracking a chunk's dense similarity score, BM25 lexical score, RRF fusion rank, and final neural reranker rank.", whyItMatters: "Reveals why a chunk was promoted or demoted at each step of the pipeline.", location: "TracePilot retrieved chunk expandable cards.", }, // ───────────────────────────────────────────────────────────────────────────── // 8. BENCHMARKING & EVALUATION (GAUGEPILOT) // ───────────────────────────────────────────────────────────────────────────── { id: "eval_faithfulness", title: "Faithfulness (Groundedness Score)", aliases: ["faithfulness", "groundedness", "hallucination score", "claim verification"], pages: ["gaugepilot", "tracepilot"], modes: ["all"], category: "Benchmarking & Evaluation", definition: "The percentage of claims in the generated answer that can be directly verified from the retrieved context (scored from 0.0 to 1.0 or 0% to 100%).", whyItMatters: "The golden metric for hallucination prevention: a score of 1.0 means every claim made by the model is strictly backed by source evidence.", formula: "Faithfulness = |Verifiable Claims in Answer| / |Total Claims in Answer|", location: "GaugePilot Leaderboard column and TracePilot evaluation score.", }, { id: "eval_relevance", title: "Answer Relevance", aliases: ["answer relevance", "relevance", "query responsiveness"], pages: ["gaugepilot", "tracepilot"], modes: ["all"], category: "Benchmarking & Evaluation", definition: "Measures how directly, concisely, and completely the generated answer addresses the user query, penalizing redundant filler or off-topic tangents.", whyItMatters: "Prevents answers that are factually true according to context but fail to answer the user's specific question.", location: "GaugePilot Leaderboards and Visualizations radar chart.", }, { id: "eval_precision", title: "Context Precision", aliases: ["context precision", "precision", "signal to noise", "top ranking"], pages: ["gaugepilot", "tracepilot"], modes: ["all"], category: "Benchmarking & Evaluation", definition: "Evaluates whether the most relevant candidate chunks are ranked at the top of the context window rather than buried at the bottom.", whyItMatters: "Essential because LLM attention is strongest at the beginning and end of the prompt ('lost in the middle' effect). High precision ensures key facts are seen first.", location: "GaugePilot Leaderboard and evaluation summary.", }, { id: "eval_recall", title: "Context Recall", pages: ["gaugepilot", "tracepilot"], modes: ["all"], category: "Benchmarking & Evaluation", definition: "Measures the proportion of ground-truth reference facts needed to answer a question that were successfully captured in the retrieved chunks.", whyItMatters: "A pipeline cannot answer questions accurately if the necessary facts were missed during the retrieval phase.", formula: "Context_Recall = |Reference Facts in Retrieved Chunks| / |Total Reference Facts in Gold Answer|", location: "GaugePilot Leaderboard and evaluation metrics.", }, { id: "eval_composite", title: "Composite Benchmark Score", aliases: ["composite score", "aggregate score", "overall score", "benchmark score"], pages: ["gaugepilot"], modes: ["all"], category: "Benchmarking & Evaluation", definition: "A weighted holistic index combining Faithfulness, Relevance, Context Precision, Context Recall, Latency score, and Cost efficiency into a single 0-100 score.", whyItMatters: "Enables straightforward ranking of competing RAG architectures on an overarching quality-versus-speed index.", location: "GaugePilot Leaderboard primary sort column.", }, { id: "eval_elo", title: "Win Rate & ELO Rating", aliases: ["elo", "win rate", "elo rating", "tournament ranking", "head to head"], pages: ["gaugepilot"], modes: ["all"], category: "Benchmarking & Evaluation", definition: "A competitive rating system that simulates head-to-head pairwise matches between competing pipeline configurations across all benchmark test cases.", whyItMatters: "Provides intuitive tournament-style ranking unaffected by scale differences between raw metric scores.", location: "GaugePilot Leaderboard 'Win Rate' and 'ELO' columns.", }, { id: "eval_pareto", title: "Pareto Frontier (Latency vs Quality)", aliases: ["pareto", "pareto frontier", "scatter plot", "tradeoff", "latency vs quality"], pages: ["gaugepilot"], modes: ["all"], category: "Benchmarking & Evaluation", definition: "The curve connecting optimal pipeline configurations where no other configuration achieves higher accuracy without increasing latency or cost.", whyItMatters: "Allows engineering teams to select the mathematically optimal architecture for their specific latency SLA (e.g., maximum quality under 800ms).", location: "GaugePilot Visualizations -> Scatter Plot & Pareto Frontier.", }, { id: "eval_radar", title: "Multi-Axis Radar Profile", aliases: ["radar chart", "radar profile", "capability polygon", "multi-metric view"], pages: ["gaugepilot"], modes: ["all"], category: "Benchmarking & Evaluation", definition: "A polygon visualization plotting a pipeline's performance across all evaluation dimensions simultaneously (Faithfulness, Relevance, Precision, Recall, Speed, Cost).", whyItMatters: "Instantly exposes strengths and trade-offs (e.g., identifying a pipeline with 99% accuracy but poor latency).", location: "GaugePilot Visualizations tab.", }, { id: "eval_ai_insights", title: "AI Analysis & Autonomous Diagnostics", aliases: ["ai analysis", "engineering recommendations", "autonomous diagnosis", "failure analysis"], pages: ["gaugepilot"], modes: ["all"], category: "Benchmarking & Evaluation", definition: "Deterministic, AI-generated architectural diagnosis analyzing benchmark failure modes and prescribing specific configuration adjustments (e.g. recommend switching from pure Dense to Hybrid + HyDE).", whyItMatters: "Automates the interpretation of complex benchmark statistics into clear, actionable engineering next steps.", location: "GaugePilot AI Analysis tab and automated diagnosis cards.", }, { id: "eval_dataset", title: "Golden Evaluation Dataset", aliases: ["golden dataset", "benchmark dataset", "ground truth", "eval questions"], pages: ["gaugepilot"], modes: ["all"], category: "Benchmarking & Evaluation", definition: "A curated collection of test queries paired with verified ground-truth reference contexts and gold-standard answers.", whyItMatters: "Ensures objective, repeatable, and statistically significant benchmarking across diverse pipeline configurations.", location: "GaugePilot Experiment Setup -> Evaluation Benchmark.", }, // ───────────────────────────────────────────────────────────────────────────── // 9. PLATFORM & ARCHITECTURE // ───────────────────────────────────────────────────────────────────────────── { id: "arch_pilotmaster", title: "PilotMaster Ecosystem", aliases: ["pilotmaster", "ecosystem", "platform"], pages: ["landing", "home"], modes: ["all"], category: "Platform & Architecture", definition: "An end-to-end Observable AI Execution Ecosystem uniting document intelligence (DocPilot), full-stack RAG telemetry & tracing (TracePilot), and automated pipeline benchmarking (GaugePilot).", whyItMatters: "Provides a unified interface where operational document interaction and rigorous experimental evaluation run against the same underlying deterministic execution pipeline.", location: "Global header and workspace switcher.", }, { id: "arch_pilotcore", title: "PilotCore Shared Execution Framework", aliases: ["pilotcore", "kernel", "execution framework", "retrieval kernel"], pages: ["landing", "home"], modes: ["all"], category: "Platform & Architecture", definition: "The shared Python kernel responsible for orchestrating document ingestion, chunking, embedding generation, vector/lexical retrieval, reranking, prompt construction, LLM streaming, and telemetry dispatch.", whyItMatters: "Ensures that queries executed during day-to-day chat produce identical deterministic results and telemetry traces as those executed in benchmark evaluation runs.", location: "Footer watermark and system architecture reports.", }, { id: "arch_prod_mode", title: "Production Mode (Standard)", aliases: ["production mode", "prod", "standard mode", "fast baseline"], pages: ["landing", "home", "docpilot", "tracepilot"], modes: ["prod"], category: "Platform & Architecture", definition: "The fast, stable, cost-effective baseline pipeline configured for day-to-day document question answering using standard dense vector retrieval and production LLM inference.", whyItMatters: "Delivers minimal latency and lower token overhead while maintaining high precision for standard knowledge retrieval tasks.", location: "Mode badge pill in header and workspace cards.", }, { id: "arch_exp_mode", title: "Experimental Mode (Research Lab)", aliases: ["experimental mode", "exp", "lab mode", "research mode", "experimentation"], pages: ["landing", "home", "docpilot", "tracepilot", "gaugepilot"], modes: ["exp"], category: "Platform & Architecture", definition: "An advanced RAG laboratory environment unlocking 11 multi-select query enhancements, dense+sparse hybrid search, multiple neural rerankers, custom chunkers, and deep DAG execution telemetry.", whyItMatters: "Allows engineers and researchers to inspect, benchmark, and compare cutting-edge retrieval strategies before rolling them into production.", location: "Mode toggle button ('🧪 Enter Experimentation Mode') and Lab badge.", }, { id: "arch_scope", title: "Conversation-Scoped Document Indexing", aliases: ["conversation scope", "scope", "document scoping", "session index"], pages: ["docpilot"], modes: ["all"], category: "Platform & Architecture", definition: "Isolating vector embeddings and search indexes strictly to the documents uploaded or selected for the active chat session.", whyItMatters: "Prevents knowledge leakage and cross-contamination between unrelated project sessions or different document sets.", location: "Top bar 'Scope: [filename]' badge and Document Selector dropdown in DocPilot.", }, { id: "arch_tiers", title: "Subscription Tiers (Free, Developer, Pro, Enterprise)", aliases: ["subscription", "tiers", "billing", "pro plan", "free plan", "token quota"], pages: ["home"], modes: ["all"], category: "Platform & Architecture", definition: "Access control levels governing monthly token allowances, active document index size limits, benchmark concurrency, and access to premium rerankers/models.", whyItMatters: "Determines the computational and model capabilities available to your account workspace.", location: "Top bar plan badge and Upgrade / Downgrade controls on Home page.", }, { id: "arch_jwt", title: "Stateless JWT Authentication", aliases: ["jwt", "auth", "token", "login", "signup"], pages: ["landing", "home"], modes: ["all"], category: "Platform & Architecture", definition: "JSON Web Tokens (JWT) signed using cryptographic keys to authenticate client requests securely without maintaining server-side session state.", whyItMatters: "Enables secure user isolation, document ownership privacy, and seamless multi-tab synchronization.", location: "Sign In / Sign Up forms and user profile pill.", }, { id: "arch_demo", title: "Quick Demo Mode", aliases: ["quick demo", "demo", "instant access"], pages: ["landing"], modes: ["all"], category: "Platform & Architecture", definition: "One-click pre-configured demo account populated with sample knowledge documents and live Groq LLM inference.", whyItMatters: "Enables immediate zero-setup exploration of the full PilotMaster suite.", location: "Landing page header '⚡ Quick Demo' button.", }, { id: "mem_working_buffer", title: "Conversational Working Memory (Session Buffer)", aliases: ["working memory", "session memory", "chat history", "conversation memory", "sliding window buffer"], pages: ["docpilot", "tracepilot"], modes: ["all"], category: "Platform & Architecture", definition: "A sliding-window buffer retaining recent conversation turns (up to 8 messages) passed directly into Query Condensation, Coreference Resolution, and generation prompts.", whyItMatters: "Allows users to ask natural follow-up questions (e.g. 'What about its pricing?') without repeating entity names or prior context.", location: "DocPilot chat sessions and TracePilot 'Working Memory' telemetry tile.", }, { id: "mem_episodic_vector", title: "Episodic Long-Term Semantic Vector Memory", aliases: ["episodic memory", "vector memory", "long term memory", "user memory index"], pages: ["docpilot", "tracepilot"], modes: ["exp", "all"], category: "Platform & Architecture", definition: "A dedicated, user-partitioned FAISS semantic memory store that indexes past discussion insights and user preferences across multiple sessions.", whyItMatters: "Enables long-term personalized recall across distinct projects and sessions without cluttering document vector indexes.", location: "PilotCore Vector Memory Engine and TracePilot Episodic Memory banner.", }, ]; /** * Intelligent Multi-Token Search & Strict Context Filter */ export function getFilteredTerms({ page = "all", mode = "all", query = "", category = "All Categories", scope = "page", // "page" | "all" }) { const normQuery = query.trim().toLowerCase(); const isExp = mode === "exp" || mode === true; // Split search query into search tokens for multi-keyword matching const searchTokens = normQuery.split(/\s+/).filter(Boolean); const matchedTerms = []; for (const term of GLOSSARY_TERMS) { // 1. STRICT Page Match (when scope === 'page') if (scope === "page" && page !== "all") { if (!term.pages.includes(page)) { continue; } } // 2. STRICT Mode Match (when scope === 'page') if (scope === "page" && mode !== "all") { if (isExp && !term.modes.includes("all") && !term.modes.includes("exp")) { continue; } if (!isExp && !term.modes.includes("all") && !term.modes.includes("prod")) { continue; } } // 3. Category Match if (category !== "All Categories" && term.category !== category) { continue; } // 4. Multi-Token Relevance Search if (searchTokens.length === 0) { matchedTerms.push({ term, score: 1 }); continue; } const titleLower = term.title.toLowerCase(); const defLower = term.definition.toLowerCase(); const whyLower = (term.whyItMatters || "").toLowerCase(); const catLower = term.category.toLowerCase(); const formLower = (term.formula || "").toLowerCase(); const locLower = (term.location || "").toLowerCase(); const aliases = (term.aliases || []).map((a) => a.toLowerCase()); let matchCount = 0; let score = 0; for (const token of searchTokens) { let tokenMatched = false; if (titleLower.includes(token)) { score += titleLower === token ? 50 : 25; tokenMatched = true; } if (aliases.some((a) => a.includes(token))) { score += 20; tokenMatched = true; } if (catLower.includes(token)) { score += 10; tokenMatched = true; } if (defLower.includes(token)) { score += 8; tokenMatched = true; } if (whyLower.includes(token)) { score += 5; tokenMatched = true; } if (formLower.includes(token) || locLower.includes(token)) { score += 4; tokenMatched = true; } if (tokenMatched) { matchCount++; } } // All search tokens must match at least somewhere for strict matching if (matchCount === searchTokens.length) { matchedTerms.push({ term, score }); } } // Sort by search relevance score descending matchedTerms.sort((a, b) => b.score - a.score); return matchedTerms.map((m) => m.term); } /** * Get count of terms specifically relevant to a given page and mode. */ export function getPageTermsCount(page, mode) { return getFilteredTerms({ page, mode, query: "", category: "All Categories", scope: "page", }).length; } /** * Friendly page names map */ export const PAGE_NAME_MAP = { landing: "Landing & Auth", home: "PilotMaster Hub", docpilot: "DocPilot Studio", tracepilot: "TracePilot Telemetry", gaugepilot: "GaugePilot Benchmark Studio", };