plenum

30 Commits

1 Branch

0 Tags

1.4 MiB

Commit Graph

Author	SHA1	Message	Date
Lasse Server	7ea92c9366	Stream the final chat answer instead of one blocking blob packages/llm/client.py:_read_stream yielded token-by-token deltas but was dead code — nothing called generate(stream=True). The chat tool loop always blocked on the final answer generation, so a ~40s answer arrived as a single SSE event and read as a hang, even though the "card stack" already streamed tool-selection progress. The final answer is structurally just whichever tool-loop iteration returns content with no tool_calls — there's no way to know in advance which iteration that'll be, and a turn can carry short narration before a tool call in the very same response. So streaming has to be speculative: - packages/llm/client.py: _read_stream becomes StreamAccumulator, which both streams live (thinking/content/tool_call) deltas and reconstructs a full ChatCompletionMessage (content, tool_calls, reasoning_content) once the stream ends. Tool-call argument fragments are accumulated by index, same pattern every OpenAI-compatible streaming client needs. Downstream tool execution code needed zero changes because of this — it just consumes the reconstructed message instead of the SDK's blocking return value. Also hardened <think>/</think> detection against a tag split across chunk boundaries (was a per-chunk substring check; now a carry-over buffer) — latent before since nothing exercised streaming for real. - backend/services/streaming_answer.py (new): AnswerStreamFilter holds buffered content back until it's implausible to still be one-line narration (this codebase already caps narration at 150 chars elsewhere), and separately holds back any trailing bracket run that could still grow into a [src:ID]/[[src:ID]] marker — citation validity isn't known until parse_and_renumber_citations runs on the complete text, so a raw marker, hallucinated or not, must never reach the client even for a frame. run_streaming_iteration drives one tool-loop iteration through this filter and emits "answer_delta" events; if a tool_call shows up after content already cleared the gate (narration grew unusually long before the model pivoted), it emits one "answer_delta_retract" so the UI doesn't strand a stale preview. - backend/services/chat.py, mp_chat.py: thread an explicit stream_answer flag through to _run_tool_loop's single generate() call site, rather than inferring streaming from event_callback being non-None — eval logging's ConversationRecorder.wrap() always returns a non-None callable, even for the plain blocking /api/chat endpoint, so that inference would have been wrong. Only stream_chat_response (the two SSE routes) sets it; /api/chat is untouched. - Frontend: both SSE consumers (ChatPanel.tsx's card stack, MpChatPanel.tsx's simpler turn list) grow "answer_delta"/"answer_delta_retract" handling. The live preview renders as plain, React-escaped text (no dangerouslySetInnerHTML, no markdown parsing mid-stream — that's a one-shot full-document transform, unsafe on partial input) and is fully replaced by the authoritative, citation-renumbered/person-linked/ language-polished HTML once the terminal "answer" event lands. Both event types were previously-unhandled and silently ignored, so the backend and frontend halves are independently safe to deploy. Added tests/test_llm_client_stream.py and tests/test_streaming_answer.py covering the reconstruction (content, tool-call fragment accumulation, split <think> tags, truncated-stream handling) and the filter (gate timing, citation-marker safety across chunk boundaries, the retract path, error propagation). 5 pre-existing failures in test_prompts_golden.py/ test_provenance.py are unrelated (reproduce identically on main).	5 days ago
Lasse Edfast	80771ff4aa	Rename the schema and code from Swedish to English talks -> speeches ("talks" reads as conference talks to everyone outside this project), motions -> documents with a new doc_type column so bills, written questions and committee reports can share the table later, and 79 columns from anforandetext -> text, intressent_id -> person_id, valkrets -> constituency, lydelse -> text, and so on. _postgres/rename_map.py is the single source of truth. The migration, its rollback, and the code rewrite are all derived from it, so they cannot drift apart. Hand- writing a rollback is how you end up with one that fails halfway through. Two identifiers could not be renamed mechanically and were done by reading the queries: dok_id means the protocol document in talks but the primary key in motions, and year stays a calendar year in speeches while becoming session_year in documents. The latter is aliased in SQL so the JSON field stays `year` and the frontend contract is unchanged. Values are never translated. Bifall and Avslag stay as published; parliament.yaml glosses them. A research tool must not silently rewrite the record. The migration is guarded by an existence check, so the same file is a no-op on a fresh database and does the work on an existing one — one schema definition in the world. ALTER TABLE ... RENAME is catalog-only, so the millions of HNSW-indexed vectors are untouched. The trigger functions are recreated explicitly, because plpgsql bodies are stored as opaque text and do not follow renames: they would have compiled fine and then failed at the next INSERT. They now read the text-search config from a database setting rather than hardcoding 'swedish'. Verified: migration round-trips to a byte-identical schema across columns, indexes and triggers; re-running is a no-op; triggers repopulate search_vector with working Swedish stemming on INSERT and UPDATE; and the renamed code runs real searches against a migrated database holding 5,000 rows of production data, with plain, prefix and exclusion syntax all working. Frontend type errors went from 9 to 7 — the rename fixed two and introduced none. Not yet done, and tracked: legacy compatibility views, the shim for SQL replayed from saved snapshots, tool-name aliases, and updating docs to the new names. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>	1 week ago
Lasse Edfast	09c1b3d79c	Initial import from rixdagen Seeded via an explicit allow-list (see /home/lasse/plenum-seed.sh) rather than by deleting files from a copy, so nothing sensitive can survive by omission. Excluded: WireGuard backup + client config, the plaintext DB password in admin.py, Arango credentials in scripts/notes.md, .claude/settings.json, a 113 MB log, providers.yaml (private endpoint), the Arango/ChromaDB-era scripts, the duplicated claude-design-system frontend copy, and assorted screenshots and one-off planning docs. 297 tracked files / 49 MB of history -> 161 files / 2.1 MB. Recovered 14 database migrations that the old .gitignore's `*.sql` rule had been hiding from version control. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>	1 week ago

Author

SHA1

Message

Date

Lasse Server

7ea92c9366

Stream the final chat answer instead of one blocking blob

packages/llm/client.py:_read_stream yielded token-by-token deltas but was
dead code — nothing called generate(stream=True). The chat tool loop always
blocked on the final answer generation, so a ~40s answer arrived as a single
SSE event and read as a hang, even though the "card stack" already streamed
tool-selection progress.

The final answer is structurally just whichever tool-loop iteration returns
content with no tool_calls — there's no way to know in advance which
iteration that'll be, and a turn can carry short narration before a tool
call in the very same response. So streaming has to be speculative:

- packages/llm/client.py: _read_stream becomes StreamAccumulator, which both
  streams live (thinking/content/tool_call) deltas *and* reconstructs a full
  ChatCompletionMessage (content, tool_calls, reasoning_content) once the
  stream ends. Tool-call argument fragments are accumulated by index, same
  pattern every OpenAI-compatible streaming client needs. Downstream tool
  execution code needed zero changes because of this — it just consumes the
  reconstructed message instead of the SDK's blocking return value. Also
  hardened <think>/</think> detection against a tag split across chunk
  boundaries (was a per-chunk substring check; now a carry-over buffer) —
  latent before since nothing exercised streaming for real.

- backend/services/streaming_answer.py (new): AnswerStreamFilter holds
  buffered content back until it's implausible to still be one-line
  narration (this codebase already caps narration at 150 chars elsewhere),
  and separately holds back any trailing bracket run that could still grow
  into a [src:ID]/[[src:ID]] marker — citation validity isn't known until
  parse_and_renumber_citations runs on the complete text, so a raw marker,
  hallucinated or not, must never reach the client even for a frame.
  run_streaming_iteration drives one tool-loop iteration through this filter
  and emits "answer_delta" events; if a tool_call shows up after content
  already cleared the gate (narration grew unusually long before the model
  pivoted), it emits one "answer_delta_retract" so the UI doesn't strand a
  stale preview.

- backend/services/chat.py, mp_chat.py: thread an explicit stream_answer
  flag through to _run_tool_loop's single generate() call site, rather than
  inferring streaming from event_callback being non-None — eval logging's
  ConversationRecorder.wrap() always returns a non-None callable, even for
  the plain blocking /api/chat endpoint, so that inference would have been
  wrong. Only stream_chat_response (the two SSE routes) sets it; /api/chat
  is untouched.

- Frontend: both SSE consumers (ChatPanel.tsx's card stack, MpChatPanel.tsx's
  simpler turn list) grow "answer_delta"/"answer_delta_retract" handling.
  The live preview renders as plain, React-escaped text (no
  dangerouslySetInnerHTML, no markdown parsing mid-stream — that's a
  one-shot full-document transform, unsafe on partial input) and is fully
  replaced by the authoritative, citation-renumbered/person-linked/
  language-polished HTML once the terminal "answer" event lands. Both event
  types were previously-unhandled and silently ignored, so the backend and
  frontend halves are independently safe to deploy.

Added tests/test_llm_client_stream.py and tests/test_streaming_answer.py
covering the reconstruction (content, tool-call fragment accumulation,
split <think> tags, truncated-stream handling) and the filter (gate timing,
citation-marker safety across chunk boundaries, the retract path, error
propagation). 5 pre-existing failures in test_prompts_golden.py/
test_provenance.py are unrelated (reproduce identically on main).

5 days ago

Lasse Edfast

80771ff4aa

Rename the schema and code from Swedish to English

talks -> speeches ("talks" reads as conference talks to everyone outside this
project), motions -> documents with a new doc_type column so bills, written
questions and committee reports can share the table later, and 79 columns from
anforandetext -> text, intressent_id -> person_id, valkrets -> constituency,
lydelse -> text, and so on.

_postgres/rename_map.py is the single source of truth. The migration, its rollback,
and the code rewrite are all derived from it, so they cannot drift apart. Hand-
writing a rollback is how you end up with one that fails halfway through.

Two identifiers could not be renamed mechanically and were done by reading the
queries: dok_id means the protocol document in talks but the primary key in
motions, and year stays a calendar year in speeches while becoming session_year in
documents. The latter is aliased in SQL so the JSON field stays `year` and the
frontend contract is unchanged.

Values are never translated. Bifall and Avslag stay as published; parliament.yaml
glosses them. A research tool must not silently rewrite the record.

The migration is guarded by an existence check, so the same file is a no-op on a
fresh database and does the work on an existing one — one schema definition in the
world. ALTER TABLE ... RENAME is catalog-only, so the millions of HNSW-indexed
vectors are untouched.

The trigger functions are recreated explicitly, because plpgsql bodies are stored
as opaque text and do not follow renames: they would have compiled fine and then
failed at the next INSERT. They now read the text-search config from a database
setting rather than hardcoding 'swedish'.

Verified: migration round-trips to a byte-identical schema across columns, indexes
and triggers; re-running is a no-op; triggers repopulate search_vector with working
Swedish stemming on INSERT and UPDATE; and the renamed code runs real searches
against a migrated database holding 5,000 rows of production data, with plain,
prefix and exclusion syntax all working. Frontend type errors went from 9 to 7 —
the rename fixed two and introduced none.

Not yet done, and tracked: legacy compatibility views, the shim for SQL replayed
from saved snapshots, tool-name aliases, and updating docs to the new names.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

1 week ago

Lasse Edfast

09c1b3d79c

Initial import from rixdagen

Seeded via an explicit allow-list (see /home/lasse/plenum-seed.sh) rather than by
deleting files from a copy, so nothing sensitive can survive by omission.

Excluded: WireGuard backup + client config, the plaintext DB password in admin.py,
Arango credentials in scripts/notes.md, .claude/settings.json, a 113 MB log,
providers.yaml (private endpoint), the Arango/ChromaDB-era scripts, the duplicated
claude-design-system frontend copy, and assorted screenshots and one-off planning docs.

297 tracked files / 49 MB of history -> 161 files / 2.1 MB.

Recovered 14 database migrations that the old .gitignore's `*.sql` rule had been
hiding from version control.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

1 week ago

3 Commits (291dcebcb8d5e3f1d9ba034b016fa0b42d018760)