কনটেন্টে যান

Reference — Glossary

Audience: anyone reading the docs who hits a term they don't recognise. Conventions: one concise definition per term, followed by a link to the canonical page where it is introduced or enforced. Terms are alphabetised. Acronyms are spelled out on first use; if you don't know one, search the page (Ctrl+F).

A

Actor The caller of a module's run() method. Either human (a person, with a name and a tenant) or service (a downstream system, with a service account and a tenant). Captured in ModuleMeta.actor for the audit log. First used in: Python API: the Workflow.run() contract

Adoption journey The four phases a deployment moves through: Discover, Pilot, Build, Scale. Each phase has its own page in the Adoption section, and each case study maps its timeline to a phase. First used in: Adoption: overview

Audit log A JSON-lines file written by the aiwf runtime. Every module call produces one row with the request id, the actor, the module version, the prompt version, the model id, the latency, the validator rejections, and the final output. The log is the source of truth for "what did the system actually do at 14:23 on Tuesday?". First used in: Reference: conventions

aiwf The short name for the deployment described in this site. The CLI is aiwf; the Python package is aiwf. Pronounced "A-I-W-F" or "ay-wiff" — both are fine. First used in: Reference: CLI

B

Banglish / Code-mixed Natural conversational text mixing Bengali and English vocabulary, transliteration, or syntax (e.g. "Net slow, router reset korlam"). Handled directly by local SLM tokenizers without an error-prone pre-translation layer. First used in: Case study: ISP support

Bar A floor, not a target. A CISO-signed-off number that a metric must meet or exceed for a release to ship. The bar moves only when the CISO moves it. First used in: Reference: benchmarks: the bar

BM25 A classic lexical retrieval method, used in the bank's RAG pipeline alongside vector search. The hybrid score is 0.5 * bm25 + 0.5 * vector. BM25 alone is good for "exact term" queries; vector alone is good for "semantic" queries; the hybrid is good for both. First used in: Case study: bank IT

Build (phase) The third phase of the adoption journey. 24/7 uptime, observability, on-call. The factory case study is the canonical Build-phase example. First used in: Adoption: build

C

CBS (Core Banking System) The core transactional banking engine (e.g. Temenos, Finacle). RAG workflows in banking must protect CBS procedures and follow strict citation boundaries. First used in: Case study: bank IT

Citation-in-set A property of an RAG response: every citation.chunk_id in the response must be in the retrieval_trace (the set of chunks the model was actually given). The validator rejects responses that fail this. 100% citation-in-set is the bank CISO's bar. First used in: Reference: prompts: RAG with citations

Containment check A substring test applied to fields of a ShiftSummary: at least 90% of the tokens in the field must appear in the input note. It is what makes the factory's summarization task a summarization task and not a generation task. The model is being asked to extract, not to invent. First used in: Reference: prompts: shift summary

D

Dead-letter queue (DLQ) A storage buffer for unprocessable or persistently failing workflow events. In local SLM workflows, payloads that fail schema or business validation twice are shunted to a DLQ for operator review. First used in: Architecture: workflow orchestration

Deterministic routing Workflow routing decisions based entirely on verified data (telemetry, SNMP alarms, database tables) rather than LLM guesses. The model only analyzes inputs when deterministic rules cannot resolve the state. First used in: Case study: ISP support

Discover (phase) The first phase of the adoption journey. The team identifies a problem, runs a feasibility study, decides if a model is the right tool. No production deployment. First used in: Adoption: discover

E

Engineer-in-the-loop The deployment pattern where a model produces a draft, a human reviews it, and the human's decision is what goes to the customer. The model is a productivity multiplier for the engineer, not a replacement. The ISP tier-1 triage is the canonical example: the model classifies; the engineer dispatches. First used in: Case study: ISP support

Eval set A frozen, hand-labelled set of cases used to grade a module. Eval sets are checked into the repo under eval/. Adding a new module requires a new eval set with a CISO-signed bar. First used in: Reference: benchmarks: the five frozen eval sets

F

FDE (Forward Deployed Engineer) An engineer embedded with a customer to ship custom AI agents on the customer's own data and infrastructure. FDEs are the role that emerged from cloud-model companies (Palantir, OpenAI, Anthropic); in local-LLM enterprises they are essential because there is no SaaS escape hatch for the data — the model has to run on the customer's hardware, the FDE has to be on the customer's site, and the handoff is the deliverable. The four phases in adoption are the FDE playbook, written in operational language. First used in: Adoption: Forward Deployed Engineering

Four-Eyes Principle A dual-authorization security control requiring two separate humans (maker and checker) before high-impact actions (e.g. database schema migrations, password resets, financial queries) can execute. First used in: Case study: bank IT

Frozen eval set An eval set that is not edited after it is merged. The cases, the gold labels, the bar — all of it is fixed. This is what makes a number reproducible. If the set is edited, the old number is no longer comparable to the new one. First used in: Reference: benchmarks: the eval set contract

G

Golden path The recommended way to deploy, including the model choice, the prompt, the validators, the eval set, and the bar. The "non-golden path" is allowed but the case studies do not document it. First used in: Adoption: overview

H

Hallucinated content In an RAG response, any claim that is not in a cited retrieved passage. The bar is 0%. The validator enforces it. A response that says "the system requires X" without a citation is a hallucination, even if X is true. First used in: Reference: prompts: RAG with citations

Hybrid search The combination of lexical keyword search (BM25) and dense vector embeddings to maximize retrieval precision across technical acronyms, account IDs, and colloquial Bengali-English questions. First used in: Case study: bank IT

L

LM Studio A local inference server. The site assumes the model runs on LM Studio, accessible at http://localhost:1234/v1. The LMStudioClient is a thin wrapper around the OpenAI-compatible /chat/completions endpoint with retry and health-check. First used in: Reference: Python API: LM Studio client

M

Mermaid Diagram-as-code syntax used throughout the documentation for flowcharts, sequence diagrams, and architecture topologies rendered directly in Markdown. First used in: Architecture: overview

Module A single-purpose wrapper around a model call. The three production modules are sla_classifier, qwen_rag, and factory_summary. Each module has its own prompt, its own validators, its own eval set, and its own ModuleMeta audit fields. First used in: Reference: Python API: the module contracts

ModuleMeta The audit fields added to every module response by the application layer. Includes request_id, actor, module_version, prompt_version, model_id, latency_ms, validator_rejections. The audit log row is built from ModuleMeta + the parsed output. First used in: Reference: Python API: the Workflow.run() contract

N

No-answer case A query where the retrieved knowledge corpus does not contain sufficient factual evidence to answer. The model must emit a deterministic fallback string rather than hallucinating an answer. First used in: Reference: prompts: RAG with citations

O

Override rate The proportion of model-generated triage suggestions rejected or altered by human operators (overrides / total). Monitored as a primary drift and degradation metric. First used in: Case study: ISP support

P

p95 The 95th percentile of a latency distribution. The ISP case study's bar is 3 s p95. "p95 under 3 s" means 95% of requests complete in 3 s; 5% can be slower. First used in: Case study: ISP support

Parallel-system coexistence The operational state where a legacy manual workflow runs concurrently with an automated AI pipeline during validation and shadow phases to eliminate operational downtime risk. First used in: Adoption: pilot

Pilot (phase) The second phase of the adoption journey. A single team, a single workflow, a single bar. The bank's IT helpdesk is the canonical Pilot-phase example. First used in: Adoption: pilot

Platinum / Gold / Silver Customer service level agreement (SLA) tiers in telecom and enterprise contracts, which deterministically drive priority assignment in triage workflows. First used in: Case study: ISP support

Priority-within-matrix The invariant rule ensuring that ticket priority (P1/P2/P3) is computed strictly by a deterministic lookup table mapping Customer Tier × Category, never by LLM speculation. First used in: Case study: ISP support

Prompt injection resistance The capability of an LLM prompt and validator pipeline to withstand adversarial instructions attempting to alter system instructions or extract sensitive data. First used in: Reference: benchmarks

Pydantic The Python validation library used for every typed object on this site. ChatRequest, ChatResult, TriageOutput, RAGOutput, ShiftSummary are all Pydantic models. The validator runs on every response. First used in: Reference: Python API: the typed contracts

Q

Qwen 2.5 1.5B Instruct The default local small language model (SLM) recommended in this site, deployed in quantized GGUF format (Q4_K_M) via LM Studio or Ollama for low-memory CPU and edge execution. First used in: Hands-on demo

R

RAG Retrieval-Augmented Generation. The pattern of retrieving relevant passages from a corpus and passing them to the model as context. The bank's helpdesk is the canonical example. First used in: Case study: bank IT

Request id A UUID assigned to each run() call. Returned to the caller, written to the audit log, attached to the user-visible response. The request id is the join key for "what did the user see, what did the model produce, and what did the validator say". First used in: Reference: Python API: the Workflow.run() contract

Retrieval trace The list of chunk_ids the model was given in the prompt, even if not cited in the response. Used to validate that every cited chunk_id is one the model actually saw. First used in: Reference: prompts: RAG with citations

ROADMAP.md The repository's permanent, date-stamped ledger recording architectural decisions, schema evolutions, and release commitments. First used in: Roadmap

S

Scale (phase) The fourth phase of the adoption journey. Multiple teams, multiple workflows, a shared platform. First used in: Adoption: scale

Schema validation The first stage of validation: the response must parse as the module's Pydantic model. A response that is valid JSON but not a TriageOutput is rejected here. First used in: Reference: Python API: the module contracts

Semver Semantic Versioning (MAJOR.MINOR.PATCH) applied strictly to typed module schemas, prompt templates, and workflow orchestration contracts. First used in: Reference: conventions

Severity (1 / 2 / 3) The categorical impact level of factory machine incidents: 1 = cosmetic / non-blocking, 2 = operational degradation, 3 = critical safety / line stoppage. Escapes are caught by safety keyword dictionaries. First used in: Case study: factory IT

sla_system The canonical Python package (aiwf.modules.sla_system) implementing ISP tier-1 complaint classification with telemetry pre-checks. First used in: Case study: ISP support

SOP (Standard Operating Procedure) Formal, version-controlled organizational procedures converted into chunked markdown documents for indexed retrieval in enterprise RAG workflows. First used in: Case study: bank IT

Stability contract The architectural guarantee that schemas, CLI arguments, and API entry points remain backward-compatible across minor versions, with changes protected by frozen eval sets. First used in: Reference: conventions

T

Temperature Sampling randomness parameter for language model decoding. Set to 0.0 across all production workflow steps to ensure deterministic and reproducible outputs. First used in: Reference: prompts

Tenant id An identifier for the customer organisation. Multi-tenant deployments route by tenant id, isolate data by tenant id, and audit by tenant id. The bank's helpdesk is single-tenant; the ISP's triage is single-tenant; the factory is single-tenant. Tenant isolation is a prerequisite for the bank and ISP deployments to grow into multi-tenant. First used in: Reference: Python API: the Workflow.run() contract

Tier × category matrix The deterministic lookup table mapping customer agreement levels (Platinum, Gold, Silver) and triage categories (Connectivity, Outage, Hardware, Billing) to service priorities (P1, P2, P3). First used in: Case study: ISP support

TOML The declarative configuration file format (aiwf.toml) used to specify engine endpoints, model timeouts, validator thresholds, and directory paths. First used in: Reference: conventions

Triage The act of routing a ticket to the right queue. The ISP's tier-1 triage is the canonical example. The model produces a TriageOutput (category + rationale + free-text confidence); the application looks up the priority in a matrix and writes the ticket. First used in: Case study: ISP support

Typed object A Pydantic model that constrains the shape and types of a value. "The response is a typed object" means "the response must parse as TriageOutput (or whichever model); anything that does not parse is rejected." The site uses typed objects as a first-class design tool. First used in: Reference: Python API: the typed contracts

U

Uptime_24h Continuous operational availability metric tracked for local workflow execution workers, targeting \(\ge 99\%\) uptime on edge infrastructure. First used in: Case study: factory IT

Usecase vs Case study An important distinction in this book: a use case is a lab tutorial or toy demonstration, whereas a case study documents real-world production deployment metrics, failure modes, and operational trade-offs. First used in: Case studies: overview

V

Validator A function that runs after the model produces a response. The site has three kinds: - Schema validator (Pydantic). "Does this parse as TriageOutput?" - Citation-in-set validator (RAG). "Are all cited chunk_ids in the retrieval trace?" - Containment-check validator (shift summary). "Are all the tokens in the field in the input?"

A response that fails any validator is rejected. The model is re-prompted once; on the second failure, the system falls back to a safe default ("please open a ticket" for RAG, category=other for triage, an empty summary for shifts). First used in: Reference: Python API: the module contracts

Vector store The embedding database (e.g. ChromaDB, Qdrant) used to store and similarity-search text chunks in enterprise retrieval workflows. First used in: Case study: bank IT

Version (module) The version of a module's code, prompt, and schema, in semver. module_version is a ModuleMeta field and an audit-log field. A breaking change to a typed object's schema is a major version bump. First used in: Reference: Python API: versioning

Version (prompt) The version of the prompt text the module is using, in the form vN. A change to the prompt text is a new file (prompts/triage.v4.txt) and a new prompt_version field. The bar is the same; the eval set is re-run. First used in: Reference: prompts: prompt versioning

W

Workflow The execution of a business process through a defined sequence of activities, validations, and human approvals. In code, represented by the Workflow.run() entry point. First used in: Reference: Python API

Workflow orchestration engine Durable execution software (Temporal, Camunda, LangGraph, Windmill) that manages state machines, retries, timeouts, and human checkpoints, calling local SLMs as stateless worker tasks. First used in: Architecture: workflow orchestration

Y

YAML The serialization format used for MkDocs site configuration (mkdocs.yml) and declarative workflow definitions. First used in: Reference: conventions

See also