Skip to content

Export pipeline

3. Export pipeline in detail

The complete single-thread export pipeline (cmd_export, export_cmd.py:42-86; the batch path cmd_batch reuses the same pipeline, batch_cmd.py:117-204). Key functions and line numbers in brackets:

flowchart TD
    START(["thread URL / UUID"]) --> IDX

    subgraph S1["① indexing (pre-step of index / batch)"]
        IDX["GraphQLClient.list_threads<br/>(graphql.py:51)<br/>LibraryThreadsRelayQuery first page 25 items<br/>LibraryRecentThreadsPaginationQuery pagination<br/>stop when endCursor is empty (graphql.py:81-85)"]
        IDX --> IDXO[("index/library_&lt;account&gt;.json<br/>cmd_index (index_cmd.py:17)")]
    end

    subgraph S2["② incremental planning (batch only)"]
        PLAN["plan_incremental pure function<br/>(hooks/incremental.py:36)<br/>sorted by lastUpdated desc, four-state classification:<br/>new / updated / done / expired"]
        PLAN --> ES{"neither full nor force?"}
        ES -->|"yes"| TRIM["trim the longest trailing done/expired run<br/>(early stop, incremental.py:83-86)"]
        ES -->|"no"| KEEP["return item by item (full-scan fallback)"]
    end

    subgraph S3["③ fetching (adapter.get_thread)"]
        GT["PerplexityAdapter.get_thread<br/>(adapter.py:58)"]
        GT --> PLAIN["ThreadFetcher.get_thread<br/>plain response (rest.py:59)<br/>GET /rest/thread/uuid<br/>entries + background_entries<br/>cursor pagination ≤20 pages, 3s between pages (rest.py:43-57)"]
        PLAIN --> DM{"detect_mode<br/>(normalize.py:66)<br/>mode detection, see §4"}
        DM -->|"computer / deep-research / council / study"| BLK["ThreadFetcher.get_thread_blocks<br/>schematized response (rest.py:67)<br/>8 SCHEMATIZED_USE_CASES (rest.py:26)<br/>sleep blocks_delay=4s before fetching (adapter.py:28,88)"]
        DM -->|"search (all signals present)"| NOBLK["skip blocks<br/>(search is simple query+answer)"]
        DM -->|"all-signals-missing fallback (adapter.py:84-87)<br/>mode=search and no entry has display_model"| BLK
        BLK --> ANOM["scan_wf_anomalies<br/>log.warning on any non-COMPLETED workflow<br/>(parsers.py:646; call site adapter.py:131-134)"]
    end

    subgraph S4["④ raw fidelity persistence"]
        RAW1[("raw_entries.json<br/>full plain dump (fs_writer.py:257-261)")]
        RAW2[("raw_blocks.json<br/>full schematized dump (fs_writer.py:262-266)<br/>absent for search threads")]
    end

    subgraph S5["⑤ parsing (parsers, schema single point)"]
        PT["parse_turn × N (parsers.py:173)<br/>steps / three-channel citation collection<br/>(entry.sources + FINAL.web_results<br/>+ WORKFLOW_ITEM_SOURCES)<br/>report_info / locked_reason into metadata"]
        EA["extract_answer (parsers.py:80)<br/>FINAL.answer JSON → plan.goals fallback"]
        ATT["attach_workflow_blocks (parsers.py:231)<br/>wf_block attached to turns by entry uuid<br/>wf_status into turn.metadata"]
        STUB["attach_stub_workflows (parsers.py:470)<br/>stub-turn 10s time-window matching"]
        UNC["collect_unconsumed_background<br/>(parsers.py:491) attribution waterfall ③"]
        SUBA["adapter.sub_agents (adapter.py:207)<br/>council nesting / background subagent mapping"]
        ATT -. "computer/council only<br/>(adapter.py:150-157)" .-> STUB
        STUB -.-> UNC
    end

    subgraph S6["⑥ rendering and persistence (fs_writer.write_thread)"]
        WJ[("thread.json (fs_writer.py:224-253)<br/>+ interruptions registration collect_interruptions (parsers.py:535)<br/>+ answer_variants registration collect_answer_variants (parsers.py:589)")]
        WM[("conversation.md compact version<br/>render_conversation (render.py:641)")]
        WT[("turns/turn_NNNN.md full version<br/>render_turn (render.py:596)")]
        WS[("sources.json + sources.md<br/>(fs_writer.py:270-278)")]
        WR[("report.md (fs_writer.py:308-316)<br/>deep-research report")]
        MF[("assets/assets_manifest.json<br/>versioned manifest (fs_writer.py:320-330)")]
    end

    subgraph S7["⑦ asset and report download"]
        ADL["adapter.get_assets (adapter.py:196)<br/>collect_downloadable_assets (parsers.py:696)<br/>same-name multi-version numbered v1..vN by created_at"]
        RDL["adapter.get_report (adapter.py:162)<br/>file_name exact match → RESEARCH_REPORT fallback<br/>skip empty drafts / binary sniffing<br/>normalize_math_delims math normalization"]
        CDN["CloudFront signed-URL direct download<br/>AssetDownloader.download_all (assets.py:80)<br/>resolve_ext three-source extension resolution (assets.py:122)"]
    end

    IDX --> S2
    S2 -->|"per-thread new/updated"| S3
    PLAIN --> RAW1
    BLK --> RAW2
    ANOM --> S5
    RAW1 -. "offline re-runnable<br/>(re-render, see offline-operations.md §12)" .-> S5
    PT --> EA
    EA --> ATT
    S5 --> S6
    RAW2 --> ADL
    ADL --> CDN
    CDN --> MF
    RDL --> CDN
    S6 --> DONE(["state.mark_ok<br/>BatchState checkpoint (state.py:123)"])

Key designs:

  1. Raw persisted first: the plain/blocks raw responses are fully written before any parsing (fs_writer.py:257-266); afterwards parsing/rendering can re-run offline — the fidelity anchor of the whole library.
  2. Plain always fetched; blocks by mode detection: search skips blocks (saves one request); when all detection signals fail, the fallback also fetches (adapter.py:74-85 comment and decision: better over-fetch than let raw_blocks.json silently go missing after a platform field change).
  3. Single parsing point: all field extraction is centralized in parsers.py (_g multi-level safe access parsers.py:23, _loads fault tolerance parsers.py:33, to_int lenient sort key parsers.py:44) — site revamps only need one place changed.
  4. Writer is read-only: the mounting of wf_block/stub_wfs/unconsumed_bgs happens in adapter.get_thread (adapter.py:150-157); FilesystemWriter doesn't mount, only consumes (fs_writer.py:217 comment).

4. Mode detection decision tree (detect_mode)

The complete decision logic of normalize.detect_mode (normalize.py:66-128). Five modes: computer / council / deep-research / study / search (search is the default).

The highest-priority signal is the platform-authoritative field entry.search_mode (SEARCH_MODE_MAP, normalize.py:50-59): official model config tested (GET /rest/models/config/v2) default_models.search=pplx_pro (UI "Best"), default_models.research=pplx_alpha (UI "Deep research"); search_mode values map one-to-one to conversation modes (archive verification: 100+ SEARCH-entry + pplx_alpha threads are 100% search_mode=RESEARCH; 100+ pure pplx_pro threads are all SEARCH). Only when search_mode is entirely absent does it fall back to the original step-name + display_model chain.

flowchart TD
    IN["Input: metadata + idx_thread + url<br/>+ turns + entries"] --> SMQ{"search_mode hit on any entry?<br/>(any over all entries, normalize.py:106-115)<br/>ASI→computer / AGENTIC_RESEARCH→council<br/>STUDY→study / RESEARCH→deep-research<br/>SEARCH/STUDIO→search"}
    SMQ -->|"hit"| MULTI{"multi-value conflict within thread?<br/>(mode switching, normalize.py:116-119)"}
    MULTI -->|"multiple values"| SPEC["take highest by specificity:<br/>computer &gt; council &gt; study<br/>&gt; deep-research &gt; search<br/>(_MODE_SPECIFICITY, normalize.py:63)<br/>+ log.warning"]
    MULTI -->|"single value"| SMCF{"conflict with downstream signals (step names/<br/>display_model)?<br/>(normalize.py:120-123)"}
    SPEC --> SMCF
    SMCF -->|"conflict"| SMWARN["log.warning recorded<br/>search_mode wins"]
    SMCF -->|"consistent or no downstream signal"| SMWIN["return the search_mode-mapped mode"]
    SMWARN --> SMWIN

    SMQ -->|"all absent"| Q1{"primary signal A: URL contains /computer/tasks/<br/>or metadata.mode == '4'<br/>or index mode ∈ ASI/COMPUTER?"}
    Q1 -->|"yes"| SM1["step_mode = computer"]
    Q1 -->|"no"| Q2{"primary signal B: a COUNCIL_RESEARCH step exists?"}
    Q2 -->|"yes"| SM2["step_mode = council"]
    Q2 -->|"no"| Q3{"primary signal C: a RESEARCH_ANSWER step exists?<br/>(content-based, not relying on Chinese labels)"}
    Q3 -->|"yes"| SM3["step_mode = deep-research"]
    Q3 -->|"no"| SM0["step_mode = (empty)"]

    SM1 --> DMS
    SM2 --> DMS
    SM3 --> DMS
    SM0 --> DMS

    subgraph DMS["fallback-chain redundant signal: display_model (DISPLAY_MODEL_MODE, normalize.py:32-37)"]
        DMQ{"display_model hit on any entry?<br/>(any over all entries, not just the first —<br/>the first entry of a mixed thread may not represent the whole)"}
        MAP["mapping table:<br/>pplx_agentic_research → council<br/>pplx_asi_opus → computer<br/>pplx_asi_opus_thinking → computer<br/>pplx_study → study"]
    end

    DMQ -->|"hit"| CF{"conflicts with step_mode?"}
    CF -->|"conflict"| WARN["log.warning records the conflict<br/>display_model wins"]
    CF -->|"consistent or step_mode empty"| DMWIN["return the display_model-mapped mode"]
    WARN --> DMWIN
    DMQ -->|"no hit"| FB{"step_mode non-empty?"}
    FB -->|"yes"| SMWIN2["return step_mode"]
    FB -->|"no"| SEARCH["return search (default)"]

Detection discipline (tested basis):

  • pplx_alpha is deliberately not in the mapping table (normalize.py:15-31 comment): it is the RESEARCH-dedicated model (default_models.research; UI fixed, no selector) — it is the target this classifier must detect, not a detection cue. The earlier statistic "121 search threads used pplx_alpha" was actually samples of this classifier's own misjudgment (without the search_mode signal, RESEARCH sessions lacking a RESEARCH_ANSWER step were judged search) — re-judged by search_mode and 124 threads migrated.
  • search_mode wins on conflict with downstream signals: it is the platform's authoritative record of the conversation mode; step names are the platform's most change-prone surface fields, display_model just a model-layer enum. Conflicts always log.warning (normalize.py:120-123).
  • Mode switching within a thread takes the highest by specificity: computer > council > study > deep-research > search (normalize.py:63, 116-119), with log.warning.
  • All-signals-missing doesn't conclude search: fall back to the pipeline-level blocks fetch (adapter.py:84-87), ensuring schematized data doesn't silently go missing due to field drift.