Answers were short, sometimes in the wrong language, and occasionally about
projects that do not exist. Measured on a 20-question battery against the real
corpus in both Spanish and English, this takes grounded content from 3/10 to
9/10 and language matching from 7/10 to 10/10.
Retrieval
- Fuse FTS5 keyword search with dense vectors via Reciprocal Rank Fusion.
Both halves are load-bearing: the corpus is English and visitors ask in
Spanish, so the meaningful words score zero. "paga" appears 0 times in a
document that says "Payments: Stripe" — the question "¿Con qué se paga en la
tienda de ropa?" retrieved nothing at all. Embeddings put all three of that
project's chunks on top. RRF ranks by agreement rather than comparing a BM25
score against a cosine, quantities with no shared scale.
- internal/embed: OpenAI-compatible embeddings client, unit-normalised so a
dot product is the cosine. Reorders by the response `index` field.
- Store a content hash beside each vector and skip rows where it no longer
matches the chunk. Chunk ids survive body edits, so without this an edited
document keeps serving embeddings that describe text that is gone —
reproduced live by changing a payment provider and watching the old one keep
coming back.
- Degrade to keyword-only when the embedder is down instead of failing.
Reference documents that are not projects
- Index `.mdx` alongside `.md`, and split sources into projects (announced in
the catalogue) and reference material (retrievable, never listed). A CV is
what someone deciding whether to hire actually reads, and it was unreachable
while it lived only in the Astro site — but filing it under projects made
the bot list "cv" as one of Victor's works.
- Skip each directory's README. `data/projects/README.md` was being indexed,
so the catalogue injected into every prompt announced "README" and
"README.es" as projects of Victor's.
- Exclude frontmatter from retrieval. It is dense metadata in a very short
chunk, which makes it a magnet for short queries: a CV's `location:` field
answered "¿Dónde ha trabajado Victor?" with a city instead of a work history.
- Split oversized sections at `###` before falling back to byte offsets. A CV's
Experience section is a list of jobs, and size-splitting cut one mid-word,
stranding the employer's name in the previous chunk.
Prompt and sampling
- Inject the full project catalogue every turn. Top-K search returns the best
matching sections, so "list every project" cannot be answered from retrieval
alone, and a small model asked to enumerate from partial hits invents the
rest. ~10 tokens per project; this is what stopped the invented names.
- Wire the sampling parameters the model authors publish (top_k, top_p, min_p,
repeat_penalty, presence_penalty) through config to llama.cpp. Leaving them
at llama.cpp's defaults produced 16-token stub answers.
- Localised system prompt selected by detected language. The English prompt
plus "reply in the user's language" answered 1/5 Spanish questions in
Spanish; few-shot examples fixed the language but got copied verbatim into
real answers.
- Fold compaction's system notes into the leading system message. Gemma's chat
template rejects a system message that is not first, and the whole request
failed with HTTP 400 the moment compaction fired.
Configuration and docs
- context_size 4096, down from 8192. The largest prompt this bot ever built
over 20 real requests was 1255 tokens, compaction starts at ~3070, and the
cut saved 212 MB resident with zero truncations and identical throughput.
- Correct the RAM figures throughout. They were measured with a GPU absorbing
llama.cpp's buffers; on a GPU-less VPS those come out of system RAM, which
is 1.1 GB more for qwen2.5-3b and 2.8 GB more for granite. Both READMEs
still started gemma-3-1b while the config defaulted to qwen, and neither
started the embedder at all.
Measured on the 2-core, 8 GB CPU-only target: 3.64 GB LLM + 0.91 GB embedder
+ 0.02 GB bot, 21.0 tok/s steady state.
Known and unfixed, so they are not re-filed as new bugs: the model reads dates
out of the CV correctly but does the arithmetic on them wrong, and "¿Dónde ha
trabajado Victor?" still answers with projects rather than employers, though
"¿En qué empresas ha trabajado?" works.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
streaming.WriteCompaction packages a 'compaction' event with the
kept/older turn counts, summary tokens and provider-reported
window/used tokens so the client can hint 'context optimized' to the
user without parsing the stream body.
streamChat runs Compact before BuildMessages and writes the event
right after start, ensuring the client sees it before any chunk is
emitted. Add a runner test that exercises limitRAGContext to keep
the system prompt + RAG block under the configured window.
Runner.Compact folds the older portion of history into a single
system-role summary when the previous turn's input tokens cross
threshold_ratio × MaxContextWindow. When summarization fails, the
runner falls back to truncateToBudget so a flaky summarize call never
breaks the user's request.
EstimatePromptTokens / totalPromptTokens give a conservative count
(roughly 3 chars per token) used by both the compaction trigger and
BuildMessages' new limitRAGContext / fitHistory helpers to cap the
prompt inside the provider's reported window before the request goes
out. Covers the first-turn case where no usage has been reported yet.
Adds runner_compaction_test.go with table-driven coverage for the
disabled, below-threshold, short-history, unknown-window, fallback
and first-stream cases, plus a regression for the RAG-context trimmer.
Adds a configurable compaction section to portfolio-bot.yaml with
threshold_ratio, keep_recent_turns and an optional summary prompt.
Wires the new fields through config.Validate() and cmd/chat-bot/main.go
into agent.Runner.WithCompaction() so the runner can opt in to
auto-compaction at startup.
Conversations survive page reloads and work for any frontend, not just
the widget. Server-side SQLite, conversation ID as bearer token, browser
identity via localStorage.
Backend
-------
- internal/portfolio/conversations.go: schema + CRUD. Conversations and
messages tables in the same SQLite DB as the RAG index, with
foreign-key cascade delete. Conv IDs are 16-byte random hex
(128 bits of entropy).
- internal/portfolio/indexer.go: applies conversation schema + enables
foreign_keys pragma in OpenStore.
- internal/server/handlers.go: POST /api/chat accepts an optional
conversation_id, mints one if absent, persists user message before
the LLM runs and assistant message (with sources) after the stream
completes. New handlers: GetConversation, ListConversations,
DeleteConversation.
- internal/server/server.go: routes for GET /api/conversations,
GET/DELETE /api/conversations/{id}.
- internal/server/conversations_test.go: 6 tests (round-trip, continue,
list, 404, delete, streaming).
Widget
------
- web/chat-widget.js: stores conv_id in localStorage["rony-chat-conv"],
includes it in the chat request body, captures new IDs from the
server's 'start' SSE event, and calls GET /api/conversations/{id} on
load to restore history. On 404 it clears the stored ID and starts
fresh.
Docs
----
- docs/architecture.md: §3.1 documents the conversation_id field and
new REST endpoints; new §3.4 covers persistence lifecycle, schema,
client responsibilities, and auth model. §5.6 updated; filetree
reflects the new files.
- web/README.md: new 'Conversation persistence' section explains the
browser-scoped behavior and how to opt out or persist across devices.
