conversation-analyser 0.2.2

Critical-thinking and analytics for human-AI conversations — a member of the lens analyser family.

conversation-analyser

Critical-thinking and analytics for human–AI conversations — a member of the analyser family.

It scores a single conversation on two tiers:

1. Analytics (always on, offline): turn/word counts, prompt/response lengths, question ratio, pushback hits, readability, sentiment trajectory, prompt self-similarity, and temporal metrics when timestamps are present.
2. Critical thinking (opt-in, needs an LLM): classifies every human turn under a 7-label prompt taxonomy, derives engagement ratios, an engagement band, and a composite 0–100 critical-thinking score with a component breakdown.

The taxonomy reuses the validated NQ/FU/CH/EX/DG/AC/MT scheme from the ISYS6020 marking pipeline (copied and forked). Design: docs/superpowers/specs/2026-05-23-conversation-analyser-design.md.

Install

pip install -e .                       # core: analytics + CLI + HTTP API
pip install -e '.[embeddings]'         # + prompt self-similarity (sentence-transformers)
pip install -e '.[llm]'                # + taxonomy/CT tier (anthropic)
pip install -e '.[embeddings,llm,dev]' # everything
export ANTHROPIC_API_KEY=...           # required for the critical-thinking tier

CLI

Bare positional path to analyse (human summary by default, --json for machines); serve subcommand for the HTTP API — same grammar as the rest of the family.

conversation-analyser transcript.txt              # human summary, analytics only
conversation-analyser chat.json --json            # full JSON to stdout
conversation-analyser chat.json --llm             # add the critical-thinking tier
conversation-analyser log.json --idle-gap 45      # split sub-sessions on 45-min gaps
conversation-analyser raw.txt --parse-mode llm-segment --llm
conversation-analyser serve --port 8009           # run the HTTP API

The critical-thinking tier is opt-in (--llm) to avoid surprise API costs; without it you get the analytics tier only.

HTTP API

conversation-analyser serve --port 8009
curl -F file=@chat.json 'http://127.0.0.1:8009/analyse'        # analytics only
curl -F file=@chat.json -F llm=true 'http://127.0.0.1:8009/analyse'
curl http://127.0.0.1:8009/health

GET /health and POST /analyse (multipart file upload, optional llm form field) — the same /analyse contract auto-analyser routes to.

Python API

from conversation_analyser import ConversationAnalyser

result = ConversationAnalyser().analyse("transcript.txt", llm=True)
print(result.model_dump_json(indent=2))

Input formats

A pluggable adapter registry tries, in order: structured adapters → heuristic speaker markers → optional LLM segmentation → unsegmented fallback.

• role/content message list (OpenAI/Anthropic): [{"role": "user", "content": "..."}, ...]
• AnythingLLM rows: [{"prompt": "...", "response": "...", "createdAt": ...}, ...]
• flat text with speaker markers: User: / Assistant: / Me: / ChatGPT: / You said: / ChatGPT said: / Prompt: / Response:
• anything else → LLM-segment (needs [llm]), else a single-blob fallback

.pdf/.docx inputs are text-extracted first (needs pdfplumber/markitdown, or pre-extract with document-analyser).

The taxonomy

Code	Meaning
NQ	New Query — opens a new topic
FU	Follow-up — clarification/elaboration
CH	Challenge — pushes back, tests, asks why
EX	Extension — applies/compares/synthesises in a new direction
DG	Delegation — task hand-off, no engagement
AC	Acknowledgement — thanks/confirmation
MT	Meta — about the conversation itself

critical_thinking = (CH+EX)/turns, delegation = DG/turns, filler = (AC+MT)/turns. Bands: One-Shot · Delegator · Directed · Iterative · Critical.

Graceful degradation

Missing	Effect
ANTHROPIC_API_KEY / [llm]	taxonomy/critical_thinking null; analytics still produced; note llm_unavailable
[embeddings]	prompt_self_similarity null; note embeddings_unavailable
timestamps	temporal metrics omitted; no sub-session split; note no timestamps

Output

ConversationAnalysis → an aggregate (rolled up over all human turns, the headline) plus one SessionAnalysis per idle-gap sub-session, each with analytics, taxonomy, critical_thinking, and per-turn turns (label + rationale + preview). See the design spec §8 for the full schema.

Testing

pytest                    # fast, deterministic (LLM mocked, no network)
pytest -m slow            # includes sentence-transformers model download
pytest -m integration     # includes live LLM calls