Short, practical definitions with a pointer to where each concept appears in this repo.
Core LLM terms
| LLM (Large Language Model) |
A neural network that predicts text; used here via the Claude API. |
llm.py |
| Prompt |
The text sent to the model: system (role & rules) + user (task & data). |
every agent’s SYSTEM |
| System prompt |
Persistent instructions defining persona, rules and output format. |
CITE_RULE, NO_FABRICATION |
| Token |
The unit of text the model reads/writes (~¾ of a word). Cost and limits are in tokens. |
max_tokens in llm.py |
| Context window |
Maximum tokens the model can consider at once (1M for Claude Opus 5). |
context_summary() keeps prompts small |
| Temperature / sampling |
Randomness control (removed on current Claude models — use effort instead). |
— |
| Effort |
How much reasoning the model spends (low…max). Cheaper for simple tasks. |
effort= per call |
| Adaptive / extended thinking |
The model reasons privately before answering. |
thinking={"type": "adaptive"} |
| Streaming |
Receive output token-by-token; avoids timeouts on long outputs. |
messages.stream(...) |
| Structured output |
Forcing the answer to match a JSON schema. |
ask_structured() |
| Pydantic model |
A Python class that validates data and generates the JSON schema. |
class Gap(BaseModel) |
| Hallucination |
Fluent but false output (e.g. invented citations). |
mitigated by tool-grounded citation rules |
| Grounding |
Tying model output to retrieved facts (papers, numbers). |
Paper.short() blocks in prompts |
| Prompt caching |
Re-using a stable prompt prefix to cut cost/latency. |
recommended for web app |
| Refusal / stop reason |
Why generation stopped (end_turn, max_tokens, refusal). |
_check() in llm.py |
Agent terms
| Agent |
An LLM that pursues a goal by taking actions (tool calls) and observing results. |
all five |
| Tool / function calling |
Letting the model call code (search, calculate, save). |
tools.py |
| Tool runner / agent loop |
The loop: model proposes tool call → code runs it → result goes back → repeat. |
pattern described in 03_how_these_agents_work.md |
| ReAct |
“Reason + Act”: interleave reasoning with tool actions. |
conceptual |
| Pipeline (workflow) |
A fixed sequence of LLM steps controlled by code. Predictable and testable. |
every main.py |
| Orchestrator |
Code (or an LLM) that decides which agent runs next. |
main.py; Research OS vision |
| Planner / worker |
One agent decomposes the task; others execute parts. |
LiteratureMiningAgent.plan |
| Critic / reviewer |
An agent that checks another’s output. |
StatisticsAgent, ReviewerSimulationAgent |
| Persona |
A role the model plays (methodologist, domain expert). |
PERSONAS in Agent 4 |
| Multi-agent system |
Several specialised agents cooperating. |
Agents 1–5 |
| Human-in-the-loop |
A person reviews/edits between steps. |
--from-evidence in Agent 2 |
| Memory (short/long-term) |
Prompt context vs. persisted state across runs. |
memory.py |
| State |
Everything the agent knows at a point in time (files, DB). |
outputs/, memory/ |
| Guardrails |
Rules/validators that constrain behaviour. |
CITE_RULE, strict schemas |
| Evaluation (evals) |
Tests measuring agent quality on known cases. |
Stage 4 of roadmap |
| LLM-as-judge |
Using a model to grade outputs against a rubric. |
evals |
| RAG (Retrieval-Augmented Generation) |
Retrieve relevant documents, then generate with them in the prompt. |
search → prompt in Agents 1–2 |
| Embedding / vector DB |
Numeric representations for semantic search. |
extension idea |
| MCP (Model Context Protocol) |
A standard for exposing tools/data to models. |
extension idea |
| Managed / hosted agents |
Provider runs the agent loop and sandbox for you. |
extension idea |
| Sandbox / code execution |
Safe environment where an agent runs generated code. |
Agent 3 extension |
| Idempotent run |
Running twice with the same input produces the same files. |
outputs/<slug>/ |
Research-domain terms
| DOI / arXiv ID / PMID |
Persistent identifiers for papers; our citation keys. |
| PRISMA |
Reporting standard for systematic reviews (identified → screened → included). |
| PICO |
Population, Intervention, Comparison, Outcome — frames a review question. |
| Inclusion / exclusion criteria |
Rules deciding which papers enter a review. |
| Evidence table |
Structured extraction: study, dataset, method, result, limitations. |
| Research gap |
A specific unanswered question or missing combination in the literature. |
| Novelty statement |
One paragraph explaining what is new and why it matters. |
| Hypothesis / research question |
Testable statement / question the study addresses. |
| Replicates, controls, covariates |
Design elements that make experiments valid. |
| QC / normalisation |
Cleaning and scaling data before analysis. |
| Differential analysis |
Testing which features differ between groups. |
| p-value / FDR (Benjamini–Hochberg) |
Significance and multiple-testing correction. |
| Cross-validation / leakage |
Honest model evaluation; leakage = test info entering training. |
| IMRaD |
Introduction, Methods, Results, and Discussion — manuscript structure. |
| Impact factor / APC / open access |
Journal metrics and publishing costs. |
| Peer review / rebuttal |
Expert evaluation and the authors’ point-by-point response. |
| Viva / defence |
Oral examination of a thesis. |
Codanics — Python ka Chilla: Build AI Agents
- Course: https://codanics.com/courses/python-ka-chilla-build-ai-agents/
- Website: https://codanics.com · YouTube: @codanics
- WhatsApp channel: https://whatsapp.com/channel/0029Va7nRDq3QxRzGqaQvS3r
- Complete free course: https://www.youtube.com/watch?v=UE3o-7KRNXs&list=PLaQmdklCcX40