Researchers and UX teams analyzing assistant conversations now face richer but messier inputs: multi-agent logs, voice calls, tool traces, and RAG context. LIA (an open-source multi-agent assistant, release 1.25.7, 19 July 2026) publishes fine-grained execution traces, HITL approval cards, voice TTS/STT pipelines, and RAG knowledge spaces. The mix is a goldmine for qualitative analysis if research teams can transcribe, align, code, and compare at scale. This post shows a practical workflow for turning LIA-style transcripts into actionable themes using AI transcription for qualitative research and how to operationalize it in Evidano (www.evidano.com). Read the source on GitHub: LIA repository on GitHub.
Key Takeaways
Evidano is an AI-powered qualitative data analysis platform that automates transcription, PII redaction, and structured ingestion for research teams. LIA 1.25.7 (19 July 2026) shows assistant data now includes execution traces, approval events, and per-message metrics, so research teams must capture meta-events as well as transcripts to produce rigorous themes. Capturing execution traces, approval cards, and observability metrics enables targeted sampling, richer analytic units, and faster triage.
- LIA release date: 19 Jul 2026, LIA v1.25.7 shows execution traces and approval cards.
- Preserve meta-events: execution traces and approval outcomes should be coded alongside message text for accountability.
- Practical pilot: export ~50 conversations with traces and run a 2-week Evidano pilot to validate coding rules and workflows.
Fast take + source
LIA 1.25.7 (19 July 2026) demonstrates what modern assistant telemetry looks like: multi-agent steps, action approvals, execution traces, and per-message cost/metric badges. For qualitative teams the richer evidence (what the assistant tried, why it failed) increases preprocessing overhead because transcripts alone miss the assistant's decisions.
- Source: LIA repo and release notes, LIA repository on GitHub
- Why it matters: logs equal context, intent, and action; transcripts alone miss the assistant's decisions
Findings snapshot, summary
This findings snapshot summarizes the key metrics and implications from the LIA release.
Findings snapshot
| Date | Metric | Value | Implication |
|---|---|---|---|
| 19 Jul 2026 | Release | LIA v1.25.7 | Execution traces & approval cards made agentic actions inspectable |
| Project stats | Code & tests | 420, 000 LOC; 11, 900+ tests; 10, 440 backend unit tests green | Mature engineering; observable behaviour backed by test evidence |
| Observability | Metrics | 423 Prometheus metrics; 25 Grafana dashboards | High signal for triage and sampling strategy |
| Scope | Agents & i18n | 19+ agents; 6 languages; RAG spaces | Multi-domain, multilingual data inputs for qualitative analysis |
What LIA brings that matters to qualitative research
LIA's output is more than chat text. For research teams LIA delivers execution traces, HITL approval cards, multimodal inputs, and per-message observability metadata:
- Execution traces: step-by-step tool calls and planner decisions that explain assistant behaviour
- HITL approval cards: explicit user confirmations and rejections recorded as coded events
- Multimodal inputs: STT/voice, attachments, images, and RAG-enriched responses
- Rich observability: per-message token and cost metadata plus hundreds of operational metrics
Consequence: research teams must capture not only utterances but also the assistant's meta-events (approved actions, reconnect banners, failed tool calls) to preserve analytic rigor.
So what for qualitative teams?
Sampling & provenance
Qualitative teams should use execution traces and approval cards as sampling keys: prioritize conversations with failed connectors, rejected approvals, or reconnect banners for deeper codebook development.
Coding richer units
Qualitative teams should treat tool calls and approval outcomes as analytic units (for example, 'planned action', 'user-approved', 'tool-failure') alongside message text to expose decision points and accountability.
Multilingual voice data
Qualitative teams should align STT transcripts with source-language audio references because LIA supports 6 languages and sentence-streamed TTS, and alignment preserves nuance for cross-language analysis.
Triage with metrics
Qualitative teams should leverage LIA's Prometheus and Grafana signals to prioritize transcripts (for example, high-cost queries, high-retry counts, unusual token patterns) and reduce wasted manual coding.
Ethics & consent (brief)
Qualitative researchers should treat assistant call summaries, voice data, and connectors as potentially sensitive, obtain consent, and redact PII before analysis.
Do more, faster with Evidano (mapping problems → features)
Problem: messy multilingual audio + tool logs
Evidano handles messy multilingual audio and tool logs by providing automated transcription and translation with custom dictionaries and PII redaction, and Evidano ingests LIA conversation exports (text, execution traces, attachments) while keeping audio links next to transcripts for verification.
Problem: inconsistent coding across analysts
Evidano addresses inconsistent coding with AI-assisted codebook import and harmonized auto-coding, allowing hierarchical codes and human-in-the-loop corrections while preserving coder edits for audit.
Problem: locating decision points in long chats
Evidano parses and indexes execution traces as structured events so teams can run thematic, content, frequency, and cross-segment analyses (for example, approved versus rejected actions) to surface patterns.
Problem: sharing concise evidence with stakeholders
Evidano produces clickable quotes, co-occurrence networks, word clouds, and exportable visual reports that combine assistant decisions with participant language in one view for stakeholder sharing.
Security & compliance
Evidano uses end-to-end encryption and does not share your data to train third-party models, which makes Evidano a fit for sensitive qualitative corpora.
Two-week workflow checklist (LIA logs → thematic report)
Use this two-week workflow to convert LIA-style data into publishable insights with Evidano:
- 1) Export: pull conversation JSON, execution traces, STT audio, and attachment files from LIA (or the hosted demo).
- 2) Transcribe & redact: run Evidano transcription with a custom dictionary for agent and tool names and automatic PII redaction.
- 3) Enrich RAG: upload LIA knowledge spaces or documents into Evidano RAG spaces so the assistant context is preserved for interpretation.
- 4) Auto-code & review: run AI-assisted auto-coding, then reconcile with a small human-coded sample to refine the codebook.
- 5) Cross-segment analysis: compare segments (approved versus rejected actions, language variants, tool-success versus tool-failure) with Evidano cross-segment reports.
- 6) Visualize & export: create co-occurrence networks and hierarchical code trees, export summary tables and clickable evidence for stakeholders.
- 7) Iterate: rerun targeted transcripts (for example, only failed-tool conversations) and run time-series or trend analyses.
FAQ: AI transcription for qualitative research
Can Evidano ingest LIA logs and execution traces?
Yes, Evidano can ingest LIA's conversation JSON and attachments. Map execution-trace fields to Evidano structured events so you can code both text and meta-events together.
How do I handle PII in voice and transcripts?
Use Evidano's PII redaction during transcription and then run a privacy QA pass on samples. Maintain exportable redaction logs for audit.
What's the best way to compare segments (for example, language, approval outcome)?
Define segments in Evidano (language, outcome tags, user role) and run thematic, frequency, and cross-segment analyses. Use co-occurrence networks to surface conditional patterns.
Wrapping up, next steps
LIA's 19 July 2026 release shows the shape of future assistant data: richly instrumented, action-capable, and multilingual, and turning that signal into research-grade themes requires transcription, structured ingestion, and cross-segment analytics. Start small by exporting 50 conversations with execution traces and run a 2-week Evidano pilot to validate coding rules. Proof the approach by prioritizing conversations flagged by observability (high-cost, failed tools, reconnect banners). Ready to try? Ingest LIA-style exports into Evidano for automated transcription, PII-safe processing, and thematic plus cross-segment analysis: Try Evidano for free. Inspect the LIA source on GitHub: LIA repository on GitHub.