| 1 | tracer provider | creates spans and tracks the current one | the agent’s process | trace and span ids, parents, start and end times, status | spans with no parent: every step its own trace |
|---|
| 2 | instrumentors | turn SDK calls into typed spans | the agent’s process | span kind, messages, token counts, tool calls | model calls as bare HTTP spans with no tokens |
|---|
| 3 | orders-api tracer and traceparent | continue the trace in a second service | orders-api | the caller’s trace id and span id | a tool span with nothing under it |
|---|
| 4 | span processor and OTLP exporter | queue finished spans and send them | every traced process | when to send, batch size, endpoint, protocol | no traces, or traces missing their last five seconds |
|---|
| 5 | OTLP receiver | accept spans | Phoenix | protobuf over HTTP on 6006, gRPC on 4317 | 401, 415 or 503 in the exporter’s log |
|---|
| 6 | database | store every span as a row | Phoenix’s SQLite file or PostgreSQL | attributes as JSON, cumulative counts, costs | an empty Phoenix after a container restart |
|---|
| 7 | UI, REST and GraphQL API | read the database for people and scripts | Phoenix, port 6006 | filters, span queries, pages of 100 to 1,000 | scripts that break after an upgrade because they read tables directly |
|---|
| 8 | annotations and evaluator job | score spans | annotation tables; a job you schedule | name, label, score, explanation, annotator kind | scores nobody can trace back to a span |
|---|
| 9 | prompt store | hold prompt versions and tags | Phoenix’s database | which version a tag points at | replies you can’t tie to a prompt version |
|---|
| 10 | datasets and experiment runner | test a change on fixed examples | Phoenix’s database; your script | examples, runs, scores | a “better” version tested on a different dataset version |
|---|
| 11 | redactor | mask personal data before export | the agent’s process | which values are masked; the customer id becomes a keyed hash | customer emails in Phoenix’s search |
|---|
| 12 | Collector | one pipeline for every service | its own process | redaction and attribute rules, where spans go | spans with names and no attributes |
|---|
| 13 | retention policy | delete old traces | Phoenix | maximum age or count, and the schedule | traces older than the privacy notice promises |
|---|
| 14 | access control | decide who reads and who sends | Phoenix | sign-in, roles, API keys | anyone on the network reading conversations |
|---|
| 15 | erasure script | forget one customer | the shop’s ops tooling | which traces and sessions go, then the copies | an erased customer back after a restore |
|---|
| · | session | group one conversation’s traces | the session.id attribute | which traces are one chat | each turn alone, with no conversation view |
|---|
| · | prompt template attributes | link a span to its prompt version | the llm.prompt_template attributes | template, variables, version | no way to filter replies by version |
|---|