docs: scope to RAG and agentic LLM apps (not voice/audio)

README.md

@@ -2,7 +2,7 @@
 
 # wiki-trace
 
-**The open-source observability platform for LLM applications.**
+**The open-source observability platform for RAG and agentic LLM applications.**
 
 Trace every request, cost, and agent step in production —
 self-hosted, zero telemetry exfiltration, one line of code.
@@ -34,9 +34,9 @@ solve this — Helicone, Phoenix, LangSmith, Weave, Datadog LLM — are
 all SaaS. Your conversations get sent to a third party.
 
 **wiki-trace is the open-source, self-hosted alternative.** Add one
-line of code and every LLM call becomes a traceable event with cost,
-tokens, latency, and replay-ability — running entirely on your
-infrastructure.
+line of code and every text LLM call in your RAG or agent pipeline
+becomes a traceable event with cost, tokens, latency, and replay-ability
+— running entirely on your infrastructure.
 
 ```python
 import wikitrace.openai
@@ -66,8 +66,8 @@ running:
 **Open-source. Apache-2.0 / MIT. Your data never leaves your machine.**
 
 > **Heads up on the name.** "wiki-trace" is a legacy from v0.1, when
-> this project was a wiki/knowledge curation tool. It's now a
-> general-purpose LLM observability platform. The name is sticking
+> this project was a wiki/knowledge curation tool. It's now an
+> observability platform for RAG and agentic LLM applications. The name is sticking
 > for now to preserve the GitHub URL and momentum — think of it as a
 > codename. The product is the tracer.
 
@@ -135,8 +135,9 @@ cd app && npm install && npm run dev
 
 ## Integrations
 
-wiki-trace is LLM-agnostic and framework-agnostic. Drop it in next to
-whatever you're already running.
+wiki-trace is provider- and framework-agnostic across text LLM APIs
+used in RAG and agent workflows. Drop it in next to whatever you're
+already running.
 
 | Provider / Framework | Integration | Status |
 |---|---|---|

ROADMAP.md

@@ -39,10 +39,10 @@ either user demand or external resources to start.
 - **Native LangSmith protocol compat** so anyone pointed at LangSmith
   can swap base URL → wikitrace, same way the Helicone-compat
   endpoints already work.
-- **Browser dashboard for streaming traces** — currently the
+- **Browser dashboard for live text traces** — currently the
   `/contribution` and `/requests` routes paint on page load. A live
   `EventSource`-fed view over `spans-live.jsonl` would close the
-  "watch your agent run" gap.
+  "watch your RAG or agent pipeline run" gap.
 - **Compliance certifications** — SOC 2 Type 1, HIPAA. Customer-driven;
   not on the path for self-hosted users.
 
@@ -52,6 +52,10 @@ either user demand or external resources to start.
   metrics to whatever ID it returns.
 - A model gateway. We use OpenRouter as the default but never proxy
   inference for billing; the proxy mode is for telemetry only.
+- Voice / audio / multimodal-realtime tracing. The span model is
+  text-completion shaped (prompt_chars, output_tokens, token deltas);
+  voice agents have an entirely different latency budget, audio
+  format, and cost model. Out of scope by design.
 - Telemetry exfiltration. Customer data never leaves their machine.
 
 ---

examples/README.md

@@ -4,9 +4,9 @@ Each script is self-contained and runnable in under 30 seconds.
 
 | Script | What it shows |
 |---|---|
-| [`openai_quickstart.py`](openai_quickstart.py) | One-line OpenAI / OpenRouter auto-patching with cost + streaming |
+| [`openai_quickstart.py`](openai_quickstart.py) | One-line OpenAI / OpenRouter auto-patching with cost + token streaming |
 | [`anthropic_quickstart.py`](anthropic_quickstart.py) | One-line Anthropic auto-patching |
-| [`decorators_quickstart.py`](decorators_quickstart.py) | `@trace` and `@tool` for arbitrary Python (sync + async) |
+| [`decorators_quickstart.py`](decorators_quickstart.py) | `@trace` and `@tool` for RAG and agent steps in Python (sync + async) |
 | [`eval_quickstart.py`](eval_quickstart.py) | Dataset, judges, `run_eval`, `compare_runs` — no LLM required |
 | [`budget_quickstart.py`](budget_quickstart.py) | Cost budgeting with `wikitrace.budget(usd=...)` |
 | [`langchain_rag.py`](langchain_rag.py) | LangChain `WikitraceCallbackHandler` end-to-end |

pyproject.toml

@@ -5,14 +5,14 @@ build-backend = "setuptools.build_meta"
 [project]
 name = "wikitrace"
 version = "0.2.0"
-description = "Open-source observability for LLM apps: trace every request, cost, and agent step. Self-hosted, MIT."
+description = "Open-source observability for RAG and agentic LLM apps: trace every request, cost, and agent step. Self-hosted, MIT."
 readme = "README.md"
 requires-python = ">=3.9"
 license = { text = "MIT" }
 authors = [
     { name = "Omkar Ray" }
 ]
-keywords = ["llm", "rag", "observability", "tracing", "evaluation", "langchain", "llamaindex"]
+keywords = ["llm", "rag", "agents", "agentic", "observability", "tracing", "evaluation", "langchain", "llamaindex"]
 classifiers = [
     "Development Status :: 3 - Alpha",
     "Intended Audience :: Developers",

sdk-js/README.md

@@ -1,7 +1,8 @@
 # wikitrace (JS / TypeScript)
 
-Local-first LLM tracing for Node and the browser. Mirrors the Python
-wikitrace SDK and emits spans to a wikitrace ingest server over HTTP.
+Local-first tracing for RAG and agentic LLM applications in Node and
+the browser. Mirrors the Python wikitrace SDK and emits spans to a
+wikitrace ingest server over HTTP.
 
 ## Install
 

sdk-js/package.json

@@ -1,7 +1,7 @@
 {
   "name": "wikitrace",
   "version": "0.2.0",
-  "description": "Local-first LLM tracing for JS/TS. Mirrors the Python wikitrace SDK and emits spans to a wikitrace ingest server.",
+  "description": "Local-first tracing for RAG and agentic LLM apps in JS/TS. Mirrors the Python wikitrace SDK and emits spans to a wikitrace ingest server.",
   "license": "MIT",
   "author": "Omkar Ray",
   "homepage": "https://github.com/OmkarRayAI/wiki-trace",
@@ -37,6 +37,8 @@
   "keywords": [
     "llm",
     "rag",
+    "agents",
+    "agentic",
     "tracing",
     "observability",
     "evaluation",